The llm package called the user-supplied onSpec/onError handlers
without awaiting them. In the /preview/stream route onSpec is async
(it does `await cacheSpec(...)` then writes the SSE `spec` event), so
the api handler's `await streamSpecFromAnthropic(...)` returned BEFORE
the terminal event had been written. The route's finally block then
ran `reply.raw.end()`, the queued `send('spec', ...)` hit a closed
stream and silently no-op'd, and the browser saw zero terminal
events — frontend ran into the "Spec generation failed." fallback
even though Anthropic had delivered a perfectly valid spec.
Verified against prod log: req-8 ran 66s with 200 and produced no
preview_spec_* log line, which is exactly the success-but-event-lost
signature.
Fix:
- StreamHandlers.onSpec / onError typed as Promise<void> | void
- Both call sites in streamSpecFromAnthropic now `await` them
- /preview/stream sets `resolved = true` at the END of each handler
(after the SSE write completes) so the post-stream "unresolved"
fallback only fires on a genuine programming bug
- Added preview_spec_ready info log on the happy path so future
diagnosis doesn't have to infer success from the absence of error
logs
@fastify/cors injects Access-Control-Allow-* in the onSend hook, but
the SSE endpoint goes straight to reply.raw.flushHeaders() — onSend
never runs, so the browser saw "No 'Access-Control-Allow-Origin'
header" and blocked the fetch before any bytes flowed.
Set Allow-Origin (reflecting the configured app origin),
Allow-Credentials, and Vary: Origin manually right before the SSE
content-type headers. Matches what the cors plugin would have
emitted on a normal response.
The DELETE /v1/servers/:id endpoint existed (tears down the runner
container + removes the row) but nothing in the UI called it, so
servers could only be removed via SSH+psql. Adds a danger-variant
button in the top-right of the detail header with a native confirm,
spinner state, and inline error surfacing. Redirects to /servers
on success.
Auth chain finally landed but tool calls crashed in the wetter server
with "Error: params is not defined". The MCP SDK passes the validated
tool args as a single parameter; our template names that parameter
`args` but the model frequently writes `params.location` / `input.x`
because that's how OpenAPI and JSON-RPC reference docs read.
Two-sided fix:
- render.ts wraps every implementation with `const params = args; const
input = args;` inside the try block. Whichever alias the model
picked, the variable resolves to the same validated object.
- SYSTEM_PROMPT now states the variable name EXPLICITLY ("variable
named EXACTLY `args`, e.g. args.location") so new generations stop
drifting on that detail.
Existing wetter runner needs a rebuild to pick up the alias shim.
Sovereign-audit Phase 3 caught the next layer of the same bug:
form-urlencoded parsing now works, but the AS metadata advertises
both `client_secret_basic` and `client_secret_post` while the handler
only read credentials from the body. Claude Desktop (and most OAuth
SDKs) prefer Basic auth, so every token exchange landed at
"401 invalid_client" — visible in prod logs as POST /oauth/token from
160.79.106.37 returning 401 in <4ms (failing the missing-secret check).
Parse Authorization: Basic header, decode base64, percent-decode each
side (RFC 6749 §2.3.1 mandates pct-encoding of user/pass before the
base64 step), and treat the resulting credentials as if they came from
the body. Header takes precedence when both are present.
Sovereign-audit traced "Authorization with the MCP server failed" past
discovery, DCR, /authorize → redirect → code, and into POST /oauth/token,
which Fastify rejected with 415 before our handler ever ran.
RFC 6749 §3.2 makes form-urlencoded the mandatory wire format for the
token endpoint, and every DCR-emitting client (Claude Desktop, Cursor,
OpenAI Codex, …) posts it that way. Fastify ships no built-in parser
for that media type so the route 415'd from the framework's content-
type layer — invisible to a code review of the route handler.
Adds a small URLSearchParams-based parser next to the existing JSON
one, parses the form body into a plain object so the route's zod
schema picks it up unchanged. No new dependency.
Architectural fix for "spec_too_large" / preview_timeout — the sync
endpoint had to fit the whole model run into Cloudflare's ~100s edge
window, which made the system fragile against any prompt that produced
a verbose spec. The new streaming path pipes Anthropic's token deltas
as Server-Sent Events; every chunk resets CF's idle timer and a 15s
keepalive comment guarantees activity even during slow first-token
windows.
@bmm/llm: new streamSpecFromAnthropic() exposes the SDK's .stream()
flow with the same typed-error contract as generateSpec — same
SpecTruncatedError / SpecValidationError / SpecTimeoutError raised from
the relevant moment.
API: POST /v1/servers/preview/stream returns text/event-stream with
events 'text' (deltas), 'spec' (final success payload, same shape as
the sync endpoint), 'error' (typed). Anthropic-only — GLM/hobby falls
back to the sync route via 409 streaming_unavailable.
Frontend: apiSseStream() handles the POST + ReadableStream + SSE
parser. The wizard's analyze() prefers the stream and only uses the
sync endpoint on the explicit 409 fallback.
nginx (api.buildmymcpserver.com): the /v1/builds/ location block (which
already had proxy_buffering off + 600s read timeout for the WS build
stream) now also matches /v1/servers/preview/stream so the SSE
response isn't buffered.
Sonnet 4.6 was still hitting max_tokens on ambitious prompts like
"WorldWeather MCP for any location" because the implementation bodies
ballooned with defensive scaffolding. Two changes:
1. SYSTEM_PROMPT now imposes hard limits the model can self-enforce:
- at most 6 tools (combine related capabilities with a mode param)
- implementation body <= 40 lines, no comments, no overengineering
- descriptions <= 100 chars
These keep a typical preview under ~7k output tokens.
2. team/enterprise maxTokens 8192 -> 12288. At ~130 tok/s that fits in
~94s, still under Cloudflare's 100s edge cap. Hobby (GLM) and pro
(Haiku) keep their existing limits — they were not hitting the
ceiling.
SpecTruncatedError still fires + surfaces 422 spec_too_large when even
12288 isn't enough, so the user gets actionable feedback instead of an
opaque zod error.
Codex/RFC review showed that Claude Desktop addresses the MCP resource
as <PUBLIC_URL>/mcp (the streamable-HTTP endpoint) rather than the
base URL. Per RFC 9728 the protected-resource metadata then lives at
.well-known/oauth-protected-resource inserted between host and path:
https://mcp.buildmymcpserver.com/.well-known/oauth-protected-resource/<slug>/mcp
Runner template now:
- publishes `resource: <PUBLIC_URL>/mcp`
- sets WWW-Authenticate to the RFC 9728 well-known URL
- serves /.well-known/oauth-protected-resource[/*] so the metadata
answers at both the legacy and RFC paths during transition
- accepts both audiences (<PUBLIC_URL>/mcp + <PUBLIC_URL>) during
rollout so already-issued tokens keep working
API:
- resolveServerByResource() tries port first, then path segment
(production path-routing), with a guard against treating "mcp" as
a tenant slug
- AS metadata advertises resource_parameter_supported: true
nginx (scripts/setup-runner-tls.sh + scripts/bmm-mcp-runners.nginx):
- new location matches /.well-known/oauth-protected-resource/<slug>/...
and proxies to the slug's runner with the slug stripped, so the
runner sees the local well-known path
Docs (oauth + api-reference) updated to the RFC paths.
Claude Desktop got past discovery + DCR but /oauth/authorize rejected
the resource parameter with invalid_resource. Root cause:
resolveServerByResource() extracted the slug from the URL's first
hostname label (subdomain routing), but production runs path routing —
mcp.buildmymcpserver.com/<slug>. The function saw resource
"https://mcp.buildmymcpserver.com/text-generation", tried to look up
slug="mcp", missed, returned null → 400.
Path lookup is now tried first (matches the production topology and
the resource URL we publish via /.well-known/oauth-protected-resource),
port lookup second (local dev), subdomain lookup last with an explicit
"mcp" guard so the legacy path doesn't shadow the new one.
Root cause of Claude Desktop's repeated "Registrierung beim
Anmeldedienst fehlgeschlagen" reference ofid_897eda676d452435:
RFC 8414 §3 constructs the well-known discovery URL by INSERTING
"/.well-known/oauth-authorization-server" between the host and the
issuer path. For issuer https://api.buildmymcpserver.com/oauth the
correct location is
https://api.buildmymcpserver.com/.well-known/oauth-authorization-server/oauth
We previously served only the issuer-appended form
(/oauth/.well-known/...), which is the historically common but
RFC-incorrect placement. Claude Desktop's MCP SDK is strict per
RFC 8414, hit the 404, and bailed out of discovery before ever
reaching /oauth/register — so the DCR fix from earlier never had
a chance to run.
Now serves the same metadata at four paths via a single handler:
- /.well-known/oauth-authorization-server/oauth (RFC 8414 strict)
- /.well-known/oauth-authorization-server (root fallback)
- /oauth/.well-known/oauth-authorization-server (historical)
- /.well-known/openid-configuration (OIDC fallback)
A single buildAsMetadata() helper keeps them in sync.
Root cause of repeat 422s: 4096 was too tight for ambitious prompts
(Marco's research-assistant prompt produces ~12kB of JSON before the
model gets cut off mid-string). The error then surfaced as an opaque
"Unterminated string in JSON" zod failure instead of pointing the user
at the real problem.
Two fixes:
- maxTokens back to 8192 (the original) for all Claude tiers, 4096 for
GLM. Timeouts bumped to 95s — Sonnet 4.6 at ~130 tok/s does 8192 in
~63s, ~30s headroom for cold starts, still under Cloudflare's 100s
edge cap.
- Detect stop_reason === 'max_tokens' on the Anthropic response BEFORE
parsing and throw the new SpecTruncatedError. /preview catches it
and returns 422 spec_too_large with a clear "split the prompt"
message instead of leaking the zod parse failure.
422s from /preview hid the actual reason: zod_message tells which field
was wrong and a 400-char preview of the model output reveals refusals
or non-JSON returns. Both stay in the api log only — never surfaced
to the client unchanged.
Enterprise plan was hitting SpecTimeoutError exactly at 60s because the
Sonnet 4.6 preview was budgeted for 8192 tokens at ~80 tok/s (≈102s
worst case) inside a 60s window. The frontend then rolled back to step
1 with no spec.
A real spec is small (<= ~10 tools, ~1.5–2.5k output tokens in practice)
so 4096 is plenty and lets even Sonnet finish in ~51s worst case. The
90s timeout buys headroom for cold starts while staying under
Cloudflare's 100s edge cap. Hobby/GLM bumped to 90s too — same
headroom argument.
Two overlapping bugs were killing OAuth discovery for every external
MCP client (Claude Desktop, Cursor, etc.):
1. worker.ts injected PUBLIC_URL=http://<RUNNER_HOST>:<port> into the
runner container even when MCP_DOMAIN was set. Result: the runner's
/.well-known/oauth-protected-resource advertised an unreachable URL
and the WWW-Authenticate header pointed at a non-HTTPS loopback
address. Claude Desktop refused to follow the discovery chain.
Now derives PUBLIC_URL from the same computePublicUrl() helper that
builds the user-visible URL stored in mcp_servers.public_url, so the
container's self-reported resource matches its actual route.
2. docker-compose.prod.yml never mounted /opt/buildmymcpserver/runner-map
into the api / generator containers. The .conf snippet written by
the generator landed in an ephemeral container path; the host
inotify watcher saw an empty directory and produced an empty
runner-map.combined. Result: nginx 404'd every /<slug>/* request,
the runner was unreachable from the public domain, and OAuth
discovery couldn't even begin. Mount added to both services.
Existing weather server has the wrong PUBLIC_URL baked in and must be
recreated after deploy. No customers yet.
export computePublicUrl from deploy.ts so worker.ts can call it.
- /oauth/register: drop resource_required check, accept generic
registrations (Claude Desktop omits resource in DCR body per spec).
serverId stored as NULL; /authorize still enforces org-ownership
+ access-token aud claim still pinned to resource. Fixes Claude
Desktop DCR failure (ofid_d7e39530c109fa7f).
- /oauth/authorize: skip strict server.id check when client.serverId
is NULL (generic client); org check remains the security boundary.
- schema: oauth_clients.server_id no longer NOT NULL.
- migration 0002: ALTER COLUMN server_id DROP NOT NULL (already
applied on prod).
- install-snippets: add Claude Code (CLI), VS Code, Codex, raw URL
tabs. Claude Desktop now shows form-field values (Name / Remote MCP
Server URL / OAuth Client ID / Secret) matching the new Custom
Connector UI instead of the obsolete JSON config.
- types: InstallTarget enum extended.
- hero-video: clicking the audio toggle restarts the video from
frame 0 so unmute aligns with the spoken opening.
- marketing: drop em-dashes from rendered copy.
Owner: "die logos müssen stimmen echte sein fetche sie." Replaced the
ASCII single-character marks (P / S / N / G / S / {}) with the actual
brand SVGs.
Sources:
- PostgreSQL, Notion, GitHub, Stripe paths from Simple Icons (CC0,
https://simpleicons.org). Inlined as React components with
fill="currentColor" so the icon colour is CSS-driven and matches
whatever foreground the brand chip uses.
- Salesforce was deindexed from Simple Icons in 2022 at the brand's
request, so I drew a clean generic cloud in the same silhouette
family — close enough to read as Salesforce-cloud-shape without
copying their trademarked mark.
- Custom REST gets a stylised pair of curly braces rendered as
stroked paths, signalling "any HTTP API" without pretending to be
a specific brand.
Brand colours used as chip backgrounds, all official values:
- PostgreSQL #336791 · Salesforce #00a1e0 · Notion #ffffff
- GitHub #181717 · Stripe #635bff · REST #6366f1
Notion is the one inversion — its mark is rendered in #0a0a0b on a
white chip because that's how Notion's actual brand mark reads. The
others all render the icon in white on a brand-colour chip.
Use of the marks is nominative fair use — they show compatibility
with each platform, not endorsement.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Owner: "die sektionen unter dem video sehen viel zu ähnlich aus — das
kannst du besser." Correct — every section was the same `panel + 3-col
grid` pattern, no page rhythm. Each section now reads as its own type
of moment:
- **Clients** ("Connects everywhere your AI lives"): typographic logo
row, no panels. Each client carries a small mono mark in a 7×7 box
(C, ⌘, ✦, <>, →) plus a 17px tracking-tight wordmark. Group hover
flips the mark and label to the accent colour so the row reads as
interactive trust signal, not a wall of text. Generous py-20/24
spacing — this is a beat between sections, not a feature card.
- **Examples** ("Wrap any HTTP API. In minutes."): asymmetric 2-col
header (h2 left, supporting copy right) over a 3-col card grid
where each integration carries a coloured 48×48 brand mark —
Postgres `#336791`, Salesforce `#00a1e0`, Notion black-on-white,
GitHub `#181717`, Stripe `#635bff`, Custom REST `#6366f1`. The marks
give each card its own visual identity, breaking the uniform-card
pattern. h2 sized 32/40 px (was a flat 28 px).
- **Marketplace** ("Skip the prompt. Fork what works."): split layout.
Left column: eyebrow + headline + supporting paragraph + bullet
list of the three selling points (no longer equal-weight cards) +
PulseLink CTA. Right column: new `MarketplaceMock` — a faux-browser
frame containing four realistic template cards (notion-search /
github-issues / stripe-readonly / linear-tasks) with author chips,
✓ verified badges, tool counts, and a fork glyph. Visitor SEES the
marketplace instead of reading copy about it.
- **Pricing** ("Pay for tool calls. Not for boilerplate."): 4-card
row but Pro is featured — indigo border, indigo glow shadow
`0 0 0 4px rgba(99,102,241,0.12)`, "RECOMMENDED" pill floating at
-top-3, and accent-coloured feature bullets. Other tiers stay
calm so the eye lands on Pro first. Price typography enlarged from
26 px to 40 px so prices read as the headline of each card.
Spacing rhythm: every section is now py-20/28 sm:py-24/28 (was
py-12-14 sm:py-16-20) — gives the below-the-fold the breathing room
it needed; the page no longer feels like a stack of crammed cards.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Owner: "wird nicht richtig gestream hab browser daten gelöscht aber kann
[nicht]" — clearing the cache didn't help. Three things changed:
1. **Single MP4 source.** Chrome listed the WebM source first because
we offered it first; on the owner's setup the VP9 decode appears to
stall silently and Chrome does NOT fall back to MP4 — it parks the
element at networkState=2/readyState=0 forever. Removing the WebM
source forces Chrome onto the MP4 (Main profile / yuv420p / TV-range
/ faststart, 2.6 MB) which we've already verified plays correctly.
2. **.load() before .play() in togglePlay.** When the original autoplay
was blocked before the source ever fetched, some Chrome builds leave
the element in a "stuck unloaded" state where subsequent .play()
calls inside a user gesture also no-op. Calling .load() first resets
the resource-selection algorithm, then .play() fetches and plays.
3. **playFailed escape hatch.** If .play() still rejects even after
.load() + user gesture (extension sandbox, hardware decoder
failure), surface a small "your browser blocked playback — open
the video directly" link to the raw MP4. The visitor isn't trapped
staring at a poster.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Owner reported "video läuft nicht, sehe nur foto" — classic blocked-
autoplay on browsers with prefers-reduced-motion / data-saver / strict
autoplay policies. The poster sat there forever and the visitor
thought the page was broken because the only control was a tiny
mute pill they didn't realise would also start playback.
Fixes:
- Tracks `playing` state via the video element's own play/pause events
so React knows whether the browser actually granted autoplay.
- Renders a large centre PLAY button overlay whenever the video is
paused. The button covers the full frame (universal YouTube / Vimeo
pattern: click anywhere on the video to play); the inner indigo
circle with the triangle is the visual affordance, with hover scale
for tactile feedback.
- Wires onClick directly on the <video> element too so the click-
anywhere-to-play works whether or not the overlay happens to be up.
- Mute toggle now calls e.stopPropagation so tapping it doesn't
accidentally trigger play/pause via the video's onClick handler.
- Best-effort .play() call in the mount effect, with the rejection
silently swallowed — failure just means the user has to click play
themselves, which the overlay already affords.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ships the long-form (71.5 s) hero video to the marketing /flow section
along with the iteration trail of architectural visual fixes the owner
worked through over the last sprint.
## Video composition (remotion/)
Eight phases driven by the 71.47 s voice-over in `audio.mp3` plus the
`Sub-bass Lullaby.wav` background music (ducked to 0.16 with fade in /
fade out). Every scene was rebuilt for v10 with concrete fixes:
- **HookScene** (12 s) — adds FloatingChaos overlay: a docker-compose
excerpt, an oauth_callback.ts snippet, an .env file with a yellow
squiggle warning ("in git history since v0.3.1"), and a live-ticking
502 retry toast. Tangle now reads as a developer's desktop right
before they give up, not as four icons drifting.
- **PromptScene** (12.2 s) — 6.5 s post-typing dead-zone replaced with
the parse beat: three sequential highlights on the prompt text
(MCP server / searches / Notion workspace), three chips below the
input (intent / tool / secret → vault), three-stat summary panel
(tools · 2, secrets · 1, targets · 3). At local frame 250 (≈ 21 s
global, on the voice line "the prompt path and the secret path
never cross") a mini two-rail diagram with an explicit X-marker
ring lands, visualising the architectural promise the moment it's
spoken.
- **SecretsScene** (15.2 s) — kept the arrow-fork + AES-256 stamp +
env-var injection beats; added the lock-snap flash at frame 66,
pinned the vault at full opacity throughout, and added a dashed
vault → container connector so the secret's provenance is visible.
The "what the AI sees" panel is now 680 px wide with an eye icon,
four corner viewfinder brackets around the prompt text, and three
explicit denied lines (no secrets / no environment variables / no
tokens).
- **BuildScene** (7.2 s) — unchanged beats: streaming log, server
card emerges with code + 🔒 NOTION_API_KEY slot pills, isolated-
container caption, <60s countdown.
- **IsolationScene** (14 s) — completely restructured. Orbit-and-dock
chips that collided with the card and with the tokens-only badge
are replaced by a clean vertical chip column at x=760: read-only
filesystem · dropped capabilities · no new privileges · 512 MB
memory cap · 0.5 CPU limit · ✓ your token only (last in green).
A vault graphic now sits below the server card with a dashed arrow
up into its env slot so the architecture story is complete in one
frame. PKCE jargon removed: "OAuth 2.1 · PKCE" → "only your token
gets in" with a small "oauth 2.1 · proof-key flow" subtitle for
the curious. Handshake stages simplified to your client → verified
→ scoped token. Final settlement arrow in success-green curves
from the scoped-token pill back into the card.
- **LibraryScene** (7 s) — cards enlarged from 340×180 to 400×220
with 36 px gaps. The "templates carry code, not credentials"
sub-caption was pulled (felt on-the-nose; the detached lock and
empty NOTION_API_KEY=? slot carry the story visually).
- **DiscoveryScene** (3 s) — the most-iterated scene. Earlier
versions had a fake "1,200+ developers building" fork counter
(pulled — solo-founder, hadn't earned). Replaced with a two-lane
architecture diagram that visualises "no paths cross" literally:
top lane prompt → AI → code, bottom lane vault → encrypted →
env, both converging at the server box on the right. v10
refinements: all seven boxes visible from frame 0 (no late
server arrival), a parallel glow tour walks across both lanes
simultaneously, a dashed vertical divider with a "no shared
node" chip pinned in the middle, and the closing line "One
sentence in. Live server out." slides down from above and lands
centred while the diagram fades to 0.12 opacity behind it —
no overlap.
- **LogoLockup** (1.7 s) — wordmark + fade-to-black for a clean
loop seam.
The Subtitle / CAPTIONS layer added in v7 was pulled wholesale —
owner found the kinetic-typography overlay aggressive and noted
that technical terms (PKCE etc.) created friction with no payoff.
Scene visuals and voice now carry the whole story; the Subtitle
component file is retained for possible future use.
Render pipeline (`render:mp4` / `render:webm` / `render:poster` in
remotion/package.json) is unchanged. The MP4 is post-processed to
H.264 Main / yuv420p / TV-range with faststart + AAC audio. The
WebM is re-encoded at VP9 CRF 38 / Opus 64k to stay under the 3 MB
budget. Final artefacts in apps/web/public/videos/: 2.59 MB mp4,
2.99 MB webm, 62 KB poster.
## Web integration (apps/web/components/hero-video.tsx)
New client component wraps the <video> element and pins a frosted-
glass mute toggle bottom-right of the player. Why not native
`controls`: the browser chrome fights the section's design vocabulary
and we only need one affordance — unmute — so we render exactly
that. The toggle's icon flips between VolumeX (currently muted) and
Volume2 (currently unmuted), accent colour switches indigo when sound
is on. Initial state is muted so autoplay still fires; on unmute we
call .play() defensively because mobile Safari pauses on
muted-property changes mid-playback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two coordinated polish moves the owner asked for.
## 1. Hero particle field — "no white dots, just a glow that follows the mouse and is always in motion"
Previous tuning (uPointSize 2.8, uBaseAlpha 0.6) gave discrete indigo
dots that additively saturated to near-white in dense clusters. The
owner wanted no granular dots visible at all — a continuous indigo
cloud that the cursor pulls toward itself.
Changes:
- **Render fragment**: replaced the anti-aliased disc SDF
(`smoothstep(0.5, 0.42, d)` — hard edge) with a Gaussian falloff
(`exp(-d * d * 6.0)` — smooth blob, no edge). Each particle is now
a soft volume that blends seamlessly with neighbours.
- **Sim fragment**: replaced the outward-gradient ring push with a
mouse-halo attraction. Particles drift toward an ideal radius
(~0.20) around the cursor, with exp-bell falloff so they don't
collapse onto the cursor or feel influenced from across the canvas.
`ringField()` helper is now unused but kept for future use.
- **JS uniforms**: `uPointSize` 2.8→14 (256-tier) / 3.6→20 (128-tier);
`uBaseAlpha` 0.6→0.055. Individual particles are below the
perception threshold for "dot" but 65k of them additively composite
into a continuous cloud. With the much lower per-particle alpha,
the cumulative brightness never saturates to white.
- **ParticleField tick loop**: asymmetric ring-active fade — `alpha
= 0.14` ramping in (fast cursor response), `0.012` decaying out
(slow glow trail after the pointer moves away). Matches the brief
"glow longer + attractive to mouse but always in motion".
- **ParticleHero index.tsx**: added an always-on indigo radial
gradient behind the WebGL canvas, so the hero never reads as
visually empty between frames — the canvas additively paints the
dynamic cloud on top. Removed the white-dot stipple from the
static fallback (it was the most likely source of the "weisse
punkte" complaint for any visitor on the fallback path).
## 2. SMS login — pre-select country picker from visitor's geo-IP
The country picker on `/login` previously defaulted to `'CH'` for
everyone. Visitors from DE / AT / US / etc. had to manually scroll
to their dial code — small friction but it sits on the highest-stakes
conversion step in the funnel.
- **New API route** `apps/api/src/routes/geo.ts` →
`GET /v1/geo/country` returns `{ country: 'CH' | 'DE' | … | null }`
by reading Cloudflare's `CF-IPCountry` header. Public, no auth —
reading a 2-letter country code from a geo-IP header isn't PII
under GDPR / DSG. `'XX'` and `'T1'` (CF's "unknown" + Tor) are
normalised to `null`. Outside CF (dev), header is missing → null.
- **Login page** picks up the result in the existing `useEffect`,
guards against codes not in our country list, and calls `setCountry`
to override the `'CH'` default. Stays at `'CH'` if the detection
fails or the visitor is on a Tor exit. Verified live: the endpoint
returns `{"country":"DE"}` from CF's German edge.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three coordinated polish items requested:
1. **Hero step-rotator tiles fit mobile without horizontal scroll.**
The previous snippets contained a 50+ char `Live at https://notion-x9.mcp.buildmymcpserver.com` URL that overflowed the ~295 px text area on a 375 px viewport. Rewrote all three snippets to be naturally short — same product story, no full URLs. The <pre> drops `overflow-x-auto` and gains `whitespace-pre-wrap break-words` so any token that does exceed the column wraps gracefully instead of forcing a scrollbar.
2. **ParticleHero — more volumetric, slower, steadier at load-in.**
The "stuttery / too fast" feedback came from two issues compounding: tiny dots (1.8 px on 256-tier, with 0.42 base alpha) gave the eye too few pixels to track between frames, so individual particles read as snapping rather than drifting; and the simplex-noise drift evolved at 0.08 time-scale with 0.045 velocity, fast enough that frame-to-frame deltas exceeded a tracked particle's diameter.
Render uniforms tuned:
- `uPointSize` 1.8 → 2.8 (256-tier), 2.4 → 3.6 (128-tier)
- `uBaseAlpha` 0.42 → 0.60
Simulation shader tuned:
- Drift noise time scale 0.08 → 0.045 (the most impactful single change — particles now move at half the previous speed)
- Drift velocity magnitude 0.045 → 0.028
- Ring breathing noise time scale 0.35 → 0.22
- Ring polar-wave time scales 1.2 / 0.7 → 0.7 / 0.42
Net effect: same number of particles (65k) but each individually larger, brighter, and moving more slowly. The cumulative additive bloom is denser without the jitter that read as visual stutter.
3. **FAQ collapsed into a native `<details>` accordion.**
Crawlers and screen readers still see every Q+A in the SSR'd HTML — `<details><summary>...</summary><p>answer</p></details>` is the standard semantic pattern for disclosure widgets. Users see one question at a time and expand on demand, which keeps the page from feeling like an endless wall of marketing text below the fold.
Container narrowed `max-w-6xl` → `max-w-3xl` for accordion typography (long-form prose reads better single-column). The default WebKit disclosure-triangle marker is suppressed with `list-none` + `[&_summary::-webkit-details-marker]:hidden`, and a `lucide-react` `ChevronDown` icon rotates 180° via `group-open:rotate-180` to indicate state.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two coordinated polish moves:
1. **<PulseLink> / <PulseButton>** — new `apps/web/components/pulse.tsx`.
Click anywhere on a wrapped link or button and a small indigo dot
detonates from the click point, scaling 1x→80x over 650ms before
fading to transparent. Same visual language as the hero load-in
glow — the click effectively says "this is the brand reaching back."
The dot lives in a `pointer-events: none` overlay, so it never
blocks the underlying navigation. `overflow-hidden + relative` are
added to the host so the bloom stays inside the rounded shape.
`glow-pulse` keyframe sits in globals.css next to the existing
`pulse-dot` / `shimmer` / `fade-in` definitions; reduced-motion
suppresses the animation to instant-opacity-0 so the click flow
is preserved without the bloom.
Wired into the highest-conversion CTAs only — the user explicitly
asked "wo's Sinn macht":
- Hero "Start building free" + "Read the docs"
- Marketing header Login / Dashboard button
- Dashboard header "+ New server" pill
Deliberately NOT applied to dashboard nav links, logout, destructive
buttons, form internals, carousel dots — pulse on every click would
be noise.
2. **Hero fills 100svh − nav** (`min-height: calc(100svh - 3rem)`).
`svh` (small viewport height) instead of `vh` so the hero doesn't
jump when the mobile address bar hides/shows. The 3rem subtracts
the sticky marketing nav (h-12 = 48px), so the hero ends right at
the loadscreen's natural bottom edge.
`flex items-center` plus the inner grid's existing `md:items-center`
keep the content vertically centred inside the tall section. The
ParticleHero background now has cinematic-scale room and the indigo
radial-glow + dot-mask read as the dominant background motif —
which is the effect the user loved at load-in.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three coordinated tweaks to the landing-page above-the-fold:
1. **Hero padding restored to py-14/sm:py-20/md:py-28** (was py-12/14/16).
Compressing it for the scroll-cue position fight made the hero feel
cramped and gave the ParticleHero background less room to breathe.
With the cue moved out (see #3), there's no reason to shrink the hero.
2. **Step rotator switches to carousel-style horizontal slide.** The
AnimatePresence transition was a fade+y-shift cross-fade — clean but
sequential. Now the leaving card slides left out (x:-220) while the
entering card slides right in (x:220→0), both coexisting in the same
3D-space and inheriting the same mouse-tilt. The container gets
`min-h-[240px]` so the absolutely-positioned cards have layout to
anchor to (claude_desktop_config.json is the tallest at 7 lines).
Reduced-motion still gets the opacity-only cross-fade — sliding
content sideways is exactly the kind of motion that preference is
meant to suppress.
3. **`<ScrollCue>` extracted into its own client component**, fixed-
positioned at viewport bottom (bottom-5) with a frosted pill style.
Fades to opacity:0 once `window.scrollY > 80`, so it doesn't shadow
the rest of the page. Lives next to `<section>` in page.tsx rather
than inside the hero — that way it anchors to the loadscreen's
natural bottom edge whether the hero is short or tall.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Restructures the landing page above-the-fold into two distinct sections:
1. **Hero — left copy + cycling tile, no static stack of three blocks**
New `<HeroStepRotator>` (Framer Motion client component) shows ONE
tile centred in the column, cycling prompt.txt → build.log →
claude_desktop_config.json every 3.5s. Auto-advance pauses on hover
and exposes a 3-dot tablist so users can jump to any step. The active
dot grows wide with an accent glow.
Mouse interaction: spring-smoothed 3D tilt on rotateX/rotateY plus a
radial glow that translates toward the cursor — both driven by motion
values, so the transforms stay on the GPU compositor instead of
re-rendering on every mousemove. `useReducedMotion()` strips the
tilt + glow translation and collapses the page transition to an
instant cross-fade (the rotation itself still advances — it's content,
not decoration).
Hero padding tightened (py-12/14/16 vs py-14/20/28) so the video
section below is teased above the fold. New scroll cue ("see it run"
+ animated chevron) sits at the bottom of the hero, anchored to
#flow.
2. **Flow video — full-width edge-to-edge under the hero (new section)**
The hero.mp4 / hero.webm pair moves out of the "How it works"
section into its own #flow section. No max-w wrapper — it spans the
viewport with `w-full aspect-video`, so on a 1080p monitor the video
gets the full 1920px width. Adds a subtle radial vignette so the
black edges blend into the page chrome.
3. **"How it works" — now lean**
Video removed (it's the flow section now). Just the three textual
cards as supporting copy.
Adds `framer-motion@11.18.2` to apps/web/package.json. Build passes
typecheck + Next.js production build with no new warnings; LCP path is
untouched since the rotator is client-hydrated after first paint and
Framer Motion is tree-shaken to the components we import.
Note: visitors with `prefers-reduced-motion: reduce` will still see the
video's poster instead of autoplay — Chrome blocks the network fetch
entirely for autoplay media when reduced-motion is set. The flow video
remains visible for the rest, and the step rotator continues to cycle
its content (with instant cross-fade instead of slide+scale).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User report: "I only see 'Search our Notion workspace' — no video."
Cause: Beat 2 (frames 55-165) was a near-empty dead moment. Particles
were 1.5-2.5px on a 1080p canvas (nearly invisible), and the server
schematic didn't start drawing until local frame 30 (= global 85),
leaving a 30-frame gap of empty space mid-clip. The viewer's brain
correctly registered "the video stops after Beat 1."
Fixes:
- 60 particles (was 36) at radius 6→3 with SVG Gaussian-blur glow
filter, always indigo (was an indecisive two-color split).
- Schematic stroke starts at local frame 8 (was 30) so the box draws
IN PARALLEL with particle convergence — eye always has something
to track.
- Central radial-glow attractor visible the whole beat — gives the
"something is forming here" cue before the schematic appears.
- Server schematic enlarged 460×300 → 720×420 so it commands
attention rather than feeling small.
- Inner tool-row dots and port dots doubled in size with stronger
drop-shadow.
- Beat 3 schematic + client panel sizes scaled to match, and the
wire base position adjusted (server CX moved from 960 to 760 so
the wire has room to breathe before reaching the client).
- Poster frame moved from 60 (mid-fade dead spot) to 180 (Beat 3
Connection layout — the most "this is a real product" shot).
File sizes still well under budget: 514 KB mp4, 319 KB webm, 29 KB poster.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New @bmm/video workspace at remotion/. Renders an 8s 1920×1080 H.264
+ WebM + JPG poster sequence that visualises the three-step "How it
works" pitch literally:
- Beat 1 (0-2s): "Search our Notion workspace" word-by-word entrance
with spring-in from below + brief indigo under-glow + monospace
prompt.txt label. Blinking cursor bridges the loop seam.
- Beat 2 (2-5s): each prompt word detonates into ~9 particles per
word; particles drift, then magnetically converge onto target slots
along a server schematic that strokes itself on. Scan-line sweep +
corner labels (mcp-notion, OAuth 2.1, search_pages, get_page_content)
sell that this is a real artefact, not a placeholder.
- Beat 3 (5-8s): Claude Desktop client panel slides in from the right;
a Bézier wire animates between server and client; three data-packet
dots travel along the wire; 200-OK tag pops; green live-dot pulses
on the server. Last 12 frames fade to black so frame 239 ≈ frame 0
and browser <video loop> has no visible seam.
Brand palette is hard-coded in lib/colors.ts to match globals.css —
keeps the Remotion bundle self-contained (no Tailwind import needed).
springIn / softSpring / clampLerp / rand helpers in lib/easings.ts
power the motion vocabulary. Concurrency=1 + yuv420p in the config
gives a deterministic render that plays on every <video> tag.
File sizes: hero.mp4 449 KB, hero.webm 258 KB, hero-poster.jpg 33 KB —
all well under the 3 MB / 250 KB ceilings.
Section 2 ("How it works") now opens with the video in a
border-bordered aspect-video panel between the heading and the three
existing cards. autoPlay+muted+loop+playsInline satisfies every mobile
autoplay policy; motion-reduce:hidden swaps in the static poster for
prefers-reduced-motion users.
Scripts:
- pnpm --filter @bmm/video render:all (mp4 + webm + poster)
- pnpm --filter @bmm/video to-web (copy to apps/web/public/videos/)
- pnpm --filter @bmm/video build (both, end-to-end)
`to-web` is the script name because `publish` collides with pnpm's
built-in npm-publish command which refused to run with an unclean tree.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds /opt/bmm-ops/ scripts (deployed separately from the app, so tar
overlays don't clobber them) for three previously-missing production
readiness items:
1. Backup hardening (backup.sh):
- Previous cron one-liner did pg_dump | gzip with no validation.
- Now: pipefail-safe pg_dump, gunzip -t integrity check, pg_dump
header sanity (scans first 5 lines — line 1 is just "--", actual
"PostgreSQL database dump" comment lands on line 2), size-warning
under 1KB, atomic move-into-place so partial backups never replace
the previous good file. 14-day retention preserved.
- Optional offsite via BMM_BACKUP_REMOTE (rclone). Reads env via
grep+cut, NOT `source` — the .env.production has unquoted text
values (e.g. ADMIN_NAME) that crash a sourced shell.
2. Restore drill (restore-test.sh, Sun 04:30 UTC weekly):
- Restores the newest backup into a throwaway DB inside the same
Postgres container, verifies the core tables exist (users,
sessions, oauth_tokens, mcp_servers), drops the temp DB. Proves
backups are actually restorable, not just byte-streams that look
like backups. Silent-corruption detector.
3. Self-hosted uptime monitor (uptime-check.sh, every 5 min):
- Probes homepage + /api/health + /robots.txt.
- Edge-triggered alerting: SMS via Twilio only on up→down and
down→up transitions (avoids SMS storm during sustained outages).
- Pings HEALTHCHECKS_HEARTBEAT_URL on every success — when the box
itself dies the heartbeat stops and the external watchdog alerts
(covers the gap that self-hosted monitors can't see their own
box failing).
notify.sh is the shared helper: Twilio SMS if all four creds set,
optional webhook to HEALTHCHECKS_FAIL_URL, always logs to syslog. Never
fails loudly — broken notification path still lands in journalctl
-t bmm-ops.
README.md documents the 3-2-1 strategy, manual full-recovery
procedure, and how to enable offsite (R2 / B2 / Hetzner Storage Box).
Smoke-tested all three on prod: backup wrote 8004 bytes with checks
passing, restore-test confirmed schema, uptime probe returned up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PageSpeed Insights mobile reported LCP element render delay of 2.3s
on the hero — the largest visible element is the build.log <pre> with
"> Generating spec... OK ..." text. TTFB is 0ms (CF cache hit), so the
delay was pure client-side: Lighthouse waited for the JS bundle to
parse and the 'use client' CodeBlock boundary to hydrate before it
considered the element "rendered."
CodeBlock pulls in lucide-react (Copy/Check icons) plus a useState
boundary just for the copy button. Above the fold on marketing, none
of that is needed — the user just needs to see the snippet.
Split:
- New `static-code-block.tsx`: server component, no 'use client',
no icons, no copy button. Pure SSR markup that paints with the HTML.
- Marketing landing now uses StaticCodeBlock for all three hero
snippets (prompt.txt / build.log / claude_desktop_config.json).
- Interactive CodeBlock stays in use for dashboard pages where users
actually want to copy snippets.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two narrow fixes for mobile chip-row width:
- Removed the 'newest' sort button. Trending and Top cover the use
cases; newest was largely redundant with Top sorted on createdAt.
- Capped the categories <select> at 140px (160px on sm+). Long
category names were stretching the box and pushing the
horizontally-scrollable chip row beyond a sane width on phones.
Native <select> truncates the visible label with ellipsis; the
dropdown panel still shows full names when opened.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two related polish items:
1. Remove the global blue Preview banner from app/layout.tsx and delete
the SiteBanner component. The component's own comment said "Remove
once the service is open for production use" — Stripe live billing,
OAuth, and per-runner TLS are all wired now, so the pre-launch notice
is misleading.
2. Mobile-responsive treatment for the standalone /templates page (it
lives outside (dashboard) layout, so it didn't inherit the new
mobile chrome from the dashboard pass):
- Top header tightened: "/templates" breadcrumb + Dashboard link +
"+ New server" pill all hidden on mobile (the avatar UserMenu +
bottom MobileActionBar cover those paths).
- Logged-in users now get the same MobileActionBar tab-bar at the
bottom (Market tab active), giving consistent app-shell across
dashboard pages.
- Filter row stacks vertically on mobile with search on top (thumb
reach), then a horizontally-scrollable chip row for scope / sort /
category so segmented controls don't squeeze below their min-width.
- h1 scales 32px → 24px on mobile; padding tightened to px-4 py-8.
- main gets pb-24 when logged in so cards clear the tab bar.
Logged-out marketplace browsing keeps the simpler marketing chrome
(Logo + "Start building" CTA) — no tab-bar, since visitors don't have
a dashboard to navigate into yet.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Top header on mobile was cramped: Logo + 5 icon-only nav buttons + avatar
crammed into a 48px-tall row. Felt like a desktop nav shrunk down.
Pivot to native-mobile-app pattern:
- Top mobile: just Logo (left) + UserMenu avatar (right). Desktop top nav
is `hidden sm:flex` so it disappears on phones.
- Bottom: full tab bar replacing the single-button MobileActionBar.
Five destinations: Overview · Servers · Create (FAB-style center) ·
Market · Settings.
- "Create" is a raised FAB-style button (round accent fill, -mt-3 to
overlap the bar border) — same prominent-action pattern as Instagram /
Notion mobile.
- Active tab gets accent color + aria-current=page.
- Audit demoted from primary nav on mobile (low frequency); still
reachable via direct /audit URL.
Desktop unchanged — top nav stays.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Same Tailwind-v4 bracket-arbitrary issue we hit on the marketing burger
menu: bg-[--color-bg-elevated] compiles to `background-color:
--color-bg-elevated` (no var() wrap → invalid color → transparent).
Both dropdowns were rendering see-through against the dashboard.
Switch both to the proven pattern: backdrop-blur-md class + inline
style for backgroundColor + borderColor using color-mix() and explicit
var(). 88% elevated-panel fill gives a clear frosted-glass look while
keeping the menu items readable.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Dashboard layout threw TypeError: Cannot read properties of null (reading
'charAt') the moment a phone-only user reached any dashboard page —
user.email and user.name are both null for fresh SMS signups, and
the initial-letter computation didn't tolerate it.
Fallback chain for the visible identifier: name → email → phone →
'Account'. Avatar colour seed falls back to userId. The secondary line
under the name also uses phone when email is null.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The per-subdomain approach (*.mcp.buildmymcpserver.com) failed at the
Cloudflare edge — Universal SSL only covers ONE-level wildcards, so the
TLS handshake on slug.mcp.buildmymcpserver.com hits SSL alert 40
handshake_failure. The two paths to fix that (CF Advanced Cert Manager
at $10/mo, or a Let's-Encrypt wildcard via DNS-01 with certbot) both
trade either money or ops for the URL aesthetic.
Pivot to path-routing on the single subdomain mcp.buildmymcpserver.com,
which IS covered by free Universal SSL. publicUrl format changes from
https://<slug>.mcp.buildmymcpserver.com → https://mcp.buildmymcpserver.com/<slug>
No recurring cost, works with the existing CF setup, MCP clients don't
care about the URL shape (it comes from the wizard's install snippet).
Code changes:
- generator/lib/deploy.ts:
* publicUrl computed as `${MCP_DOMAIN}/${slug}` instead of `${slug}.${MCP_DOMAIN}`
* writeRunnerMapEntry writes one-line nginx snippet:
if ($bmm_slug = "<slug>") { set $bmm_port <port>; }
(was: a map-entry pair "<slug>.<MCP_DOMAIN> <port>;")
- setup-runner-tls.sh:
* nginx vhost is now single server_name mcp.buildmymcpserver.com
* regex location captures (?<bmm_slug>...)(?<bmm_path>/.*)?
* includes runner-map.combined inside the location block so the
generated if-snippets set $bmm_port; unknown slug → 404
* proxy_pass strips the slug prefix: /<slug>/foo → 127.0.0.1:port/foo
* Prereq docs updated: just A-record for mcp (no wildcard needed),
same Origin CA cert reused
* Added /health endpoint at vhost root for monitoring
Systemd watcher + map dir + volume mounts unchanged — same file paths,
just different snippet content. Re-running setup-runner-tls.sh on the
host overwrites the wildcard vhost with the new path-based one.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
OAUTH REFRESH-TOKEN
- oauth_tokens.subject column added (migration applied to prod DB): stores
the JWT sub claim from the original authorization so refreshes can
re-mint with the same identity without re-walking the (consumed) code.
- Authorization-code branch now writes subject AND uses a 30-day
expires_at for the row (was 1h — same as access token, which killed
refresh after 1h).
- New refresh_token grant branch:
* looks up token by refresh-hash + expiry
* client_id must match, client_secret verified if confidential
* RFC 8707: requested resource must equal stored resource
* OAuth 2.1 rotation: atomic UPDATE WHERE old_hash → new access JWT,
new refresh token, extended expiry; loser of a race sees invalid_grant
- Access TTL (1h) and refresh TTL (30d) extracted as constants.
Clients no longer have to re-authorize hourly. Closes Zb-001.
PER-RUNNER SUBDOMAIN TLS (Z1-002)
Code path:
- New MCP_DOMAIN env (e.g. "mcp.buildmymcpserver.com") + RUNNER_MAP_DIR
(default /var/runner-map) in generator config.
- deployContainer: writes /var/runner-map/<slug>.conf with content
"slug.MCP_DOMAIN port;" and computes publicUrl as
https://<slug>.<MCP_DOMAIN>. Falls back to http://host:port when
MCP_DOMAIN is unset (zero behaviour change until host is configured).
- stopContainer (both api/lib/docker.ts and generator/lib/deploy.ts) now
accepts an optional slug arg and removes the map fragment. Callers
(DELETE /v1/servers/:id, admin template takedown) updated.
Infra path (one-time host setup — Marco runs as root):
- scripts/setup-runner-tls.sh:
1. nginx vhost matching *.mcp.buildmymcpserver.com via regex →
reads slug→port from /opt/buildmymcpserver/runner-map.combined
2. systemd inotify service watches the map dir, combines fragments
on any change, reloads nginx
3. installs inotify-tools if missing, idempotent
- Prereqs documented at top: Cloudflare wildcard DNS proxied, Origin CA
cert for *.mcp.buildmymcpserver.com, SSL mode Full (strict).
- After running: edit docker-compose.prod.yml to mount the map dir into
api + generator, set MCP_DOMAIN in env, recreate containers.
Closes Zb-001 fully. Closes Z1-002 on the code side; one Marco-on-host
action away from closing it on the infra side.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Native <select> defers dropdown direction to the browser, which on mobile
routinely opens upward and hides countries behind the keyboard. Replaced
with a custom combobox that always opens DOWNWARD (absolute positioned
below the trigger) with a search input at top — at 150 countries a
scrollable list is unusable without search anyway.
COUNTRIES list expanded from 60 → 152 entries: every country with a
meaningful diaspora, including Russia, Pakistan, Bangladesh, Sri Lanka,
Cyprus, Malta, Albania, Bosnia, Kosovo, North Macedonia, Iran, Iraq,
Lebanon, Jordan, Kazakhstan, Morocco, Algeria, Tunisia, Ethiopia,
Tanzania, Uganda, Senegal, Ghana, Madagascar, Cameroon, Sri Lanka,
Belarus, Georgia, Armenia, Azerbaijan and the rest. Serbia was already in
the prior list — just unfindable without search.
Bonus: flag emojis computed from ISO-3166 alpha-2 codes (no asset files).
Search matches name + code + dial-prefix so "+41" or "CH" both find
Switzerland.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The clearCookie call on /v1/auth/logout was passing only {path:'/'},
missing the httpOnly + sameSite + secure flags the setCookie used. In
production (secure=true), Chrome treats a Set-Cookie clear directive
without Secure as a *different* cookie — it creates an empty insecure
cookie and leaves the original Secure session cookie in place. Result:
users who clicked "Sign out" stayed logged in for the full 30-day
session lifetime in the browser's view (DB session was destroyed
correctly; only the cookie persisted).
Now both setCookie and clearCookie pull from sessionCookieOpts() so
the attributes can't drift apart again.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Logged-out state was showing two CTAs ("Sign in" link + "Start building"
button) both going to /login — confusing because the prominent purple
button never literally said "Login". Consolidate to one button whose
label flips with auth state: "Login" when out, "Dashboard" when in.
Same slot, same colour, no header layout shift.
Defaults to "Login" while the /v1/auth/me probe is in flight so the
common (anonymous) visitor sees no flicker.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the dependency on an unbuilt email sender. New EMAIL_AUTH_ENABLED
env flag (default false). When off:
- POST /v1/auth/magic-link → 503 email_auth_disabled
- POST /v1/auth/verify → 503 email_auth_disabled
- GET /v1/auth/providers → { email: false, sms, google, github }
- Login page: hides the email/phone tab toggle (only one method),
hides the email form entirely, defaults to SMS/phone tab
Flipping EMAIL_AUTH_ENABLED=true re-enables the magic-link routes and
re-shows the email form section. Schema (magic_links table) unchanged
so this is a 1-env-flip re-enable, not a re-implementation.
SECURITY: closes audit finding Za-001 (account-takeover via
cross-provider email lookup). Without a magic-link flow, an attacker
who controls a target's inbox can no longer claim an existing
OAuth-created account. The remaining provider-mixing surface (Google
↔ GitHub at same email) requires controlling the OAuth provider
account itself, which is each provider's own security boundary.
Active login methods now: Google OAuth · GitHub OAuth · SMS code
(Twilio) · admin password (seeded, single user).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Six confirmed findings closed (3 MEDIUM, 3 LOW). Tier-1 surfaces from
Pass-1 re-verified non-regressed; this pass deepened the audit on the
auth library, OAuth issuer, and template marketplace.
Za-002 MEDIUM (scrypt cost) — bump SCRYPT_N from 2^14 → 2^17 (131072)
matching current OWASP guidance for password hashing in 2026. Hash
format embeds N (`scrypt$N$salt$hash`), so the existing admin
password at the old cost still verifies — backward-compatible. Also
added explicit maxmem ceilings since Node's default (~32MiB) is
insufficient for the new N.
Za-003 MEDIUM (single-use race) — consumeMagicLink was SELECT-then-
UPDATE; two parallel redemptions could both win and mint two
sessions from the same token. Now uses the same atomic
`UPDATE … WHERE id = ? AND consumedAt IS NULL RETURNING id` pattern
/oauth/token already had — loser of the race gets
invalid_or_expired_token.
Za-004 LOW (membership ordering) — `.orderBy(memberships.createdAt)`
added so when org-invites eventually let a user belong to multiple
orgs, the same one wins every login instead of insertion-order
roulette. Latent-bug pre-empt.
Zb-002 LOW (OAuth register spam) — /oauth/register now per-IP daily
rate-limited at 20/day (well above any legitimate MCP-client
bootstrap pattern). Prevents DB-row spam.
Zc-001 MEDIUM (banned-pattern drift) — three separate copies of
BANNED_PATTERNS had drifted apart. The publish-time scanner in
templates.ts was MISSING the 7 new patterns added in Pass-1
(process.binding, dlopen, .constructor.constructor, vm.runIn*,
globalThis['..']). Single source of truth in @bmm/llm now exports
SHARED_BANNED_PATTERNS; templates.ts composes PUBLISH_BANNED_PATTERNS
= SHARED ∪ code-only-extras (dynamic import, fs.rm, setTimeout-with-
string, process.kill, jailbreak markers).
Zc-002 LOW (N+1) — /v1/templates list was issuing one COUNT(*) per
template (101 queries for a 100-row page). Now one grouped query
with templateId GROUP BY, merged in JS. p95 doesn't degrade with
marketplace growth.
DEFERRED (documented, scoped for next sprint):
Za-001 HIGH — Account takeover via cross-provider email lookup.
Requires schema change (users.primaryProvider). Mitigation in
/settings/account banner planned.
Zb-001 MEDIUM — /oauth/token refresh_token grant: advertised in
AS metadata but unsupported_grant_type. Either implement (~40
LOC) or strip from metadata.
Zc-003 LOW — Admin takedown partial-failure consistency.
Zd-001 IMPROVE — DEK cache invalidation across replicas (single-
instance today).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five confirmed findings from the sovereign-audit pass, ordered by severity:
Z3-001 CRITICAL — Fastify now trustProxy:true so req.ip resolves to the
real visitor IP via X-Forwarded-For instead of always being the nginx /
docker-bridge peer. Every per-IP rate-limit in the codebase was silently
collapsed into one global counter; this restores them.
Z1-001 CRITICAL — runner container hardening flags (--read-only,
--cap-drop=ALL, --security-opt=no-new-privileges:true, --pids-limit=100,
--memory=512m, --cpus=0.5, tmpfs /tmp) were sitting commented-out as a
TODO despite /security promising them. Now applied unconditionally on
production/staging; opt-out flag RUNNER_DISABLE_HARDENING=1 for Win-dev.
Z2-001 + Z2-002 CRITICAL / MEDIUM — banned-pattern blacklist tightened
(Function(...) without `new`, process.binding, process.dlopen,
.constructor.constructor, _load, vm.runIn*Context, globalThis['..'],
"system prompt override"). scanForInjection now also walks tool.name and
every inputSchema property description, not only implementation +
description — closes the prompt-injection-into-AI-client surface that
downstream clients (Claude Desktop, Cursor) read verbatim. The duplicate
BANNED_PATTERNS in apps/api/src/routes/servers.ts deleted in favour of
the single shared scanForInjection export from @bmm/llm.
Z4-001 HIGH — /v1/auth/magic-link gained the two-axis daily rate-limit
the SMS endpoint already had: 10/IP/day + 5/email/day. Combined with the
trustProxy fix above these are now real per-visitor limits.
Z4-002 MEDIUM — magic-link callback URL no longer printed to stdout in
production. In dev it still prints (so devs can click the link); in
production we log only "issued, URL withheld" and a loud error if no
email sender is wired (Resend integration is the actual launch
blocker — left as a TODO).
Z6-001 MEDIUM — /v1/builds/:id/stream WebSocket now refuses cross-origin
upgrades. SameSite=Lax already mitigates in modern browsers; this is the
defense-in-depth against browser bugs and non-browser clients.
FALSE POSITIVES dismissed: slug path-traversal (schema regex
^[a-z][a-z0-9-]*$ in @bmm/types catches it); session-after-promote
(getSession re-fetches isAdmin from DB on every request).
DEFERRED (not blockers, tracked):
- Z1-002 generated-server HTTPS — needs nginx wildcard subdomain TLS
- Z1-003 docker image cleanup cron
- Z2-001 v2 — real sandbox runtime (multi-week refactor)
- Z3-002 rawBody-per-request memory — branch on webhook path only
- Z5-001 multi-user org RBAC for billing — gated on Team feature
- Email sender integration (Resend) — launch blocker
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User-facing identity:
- UserMenu component in dashboard header: avatar (deterministic colour from
email hash), email + name, current plan badge, dropdown to Profile /
Billing / Support / Your data / (Admin panel if isAdmin) / Sign out
- /settings/profile: editable display name; email + phone shown read-only
(changing them requires support ticket — magic-link flow assumed)
- GET + PATCH /v1/account/profile
In-app subscription management (no more Stripe Portal redirect for the
common flows — cancellation, plan switch, invoice viewing all in-app):
- Billing status now combines DB state with a live Stripe lookup of the
subscription details + last 5 invoices. Single roundtrip.
- POST /v1/billing/cancel → schedules cancel_at_period_end
- POST /v1/billing/reactivate → undo scheduled cancel
- POST /v1/billing/change-plan → prorated swap between any tier+cycle
- /settings/billing rewritten: current plan card with renew/cancel date,
big cancel button + reactivate flow, plan-switcher grid, invoice list with
PDF + hosted-invoice links
- Stripe portal still linked at the bottom as the escape hatch for rare
actions (payment-method update, address change). New-subscription Checkout
still uses Stripe-hosted Checkout (industry standard for PCI).
Stripe SDK v22 / API 2024-09 fix: current_period_end moved to subscription
items; updated read paths accordingly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Table-cell Link only wrapped the subject text — clicks on email/status/time
cells did nothing, which read as 'cannot open ticket' for the admin. Convert
to a flex-grid Link wrapping the entire row, same pattern as the user-side
/settings/support list.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The /admin/support page existed but was invisible from the panel — sidebar
NAV array didn't list it. Adds Support as the 2nd nav item (right after
Overview, since unanswered tickets are the most-time-sensitive thing an
admin checks). Sidebar polls /v1/admin/support/counts every 30s and renders
an amber count badge next to the entry when tickets are awaiting_admin.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- @bmm/api: stripe@22 SDK, plan-aware price-id lookup, Redis-backed event
idempotency (7d TTL covers Stripe's retry window), startup warning when
STRIPE_PRICE_* env vars contain product ids (prod_) by mistake
- routes/billing.ts:
POST /v1/billing/checkout-session → Stripe-hosted Checkout, SEPA+card,
auto-VAT via Stripe Tax, tax_id
collection for B2B, address required
POST /v1/billing/portal → Customer Portal session
GET /v1/billing/status → drives the settings/billing UI
POST /v1/billing/webhook → signed, idempotent, handles
checkout.session.completed,
subscription.{created,updated,deleted},
invoice.{paid,payment_failed}
- index.ts: rawBody-aware JSON parser so Stripe signature verify gets the
exact payload bytes
- web: /settings/billing page (status, upgrade flow, manage-billing portal,
auto-checkout when arriving with ?tier=… from the pricing CTAs), pricing
page CTAs point to /settings/billing?tier=…
- Payment-failure path: suspend org only after 3rd failed attempt (Stripe
Smart Retries handles the soft-retries). Suspended orgs keep their running
servers but cannot create new ones (enforcement is in /v1/servers POST as
a follow-up).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The earlier caps (Team 150/day, Enterprise 1000/day) used Sonnet/Opus pricing
that put max-usage above the tier's monthly revenue — a Bot with a Team
subscription could out-cost €199 in Anthropic spend. Drop to 50/day Team
and 200/day Enterprise; both now keep ~55-65% margin even when maxed.
Pricing page Team feature line updated to match (150 -> 50). Build caps
loosened slightly less since the 24h cache TTL makes most builds cache-hits.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The free tier was hemorrhaging Anthropic cost with no abuse cap (no rate
limit on /preview, Opus default in the build worker, 5-min cache TTL that
made cache-miss the common case). This switches free users to GLM, paid
users to Claude tiers, and tightens every leak found in the audit.
Backend:
- @bmm/llm: GLM provider via Zhipu's OpenAI-compatible endpoint, pickPreviewModel
+ pickBuildModel helpers, plan-aware ModelChoice
- preview-cache TTL 5min -> 24h (kills the cache-miss path)
- /v1/servers/preview: picks model from caller's plan, returns model name to UI
- /v1/servers POST: enforces SERVER_LIMITS per plan (402), rate-limits builds
- daily rate-limit on preview (5/40/150/1000) and build (3/20/100/500)
- /v1/auth/me returns plan so the wizard can show the right model name
- generator worker: GLM default, Anthropic Sonnet fallback if GLM errors
Frontend:
- Wizard fetches plan, shows "<model> is drafting the tool spec" pre-emptively,
upgrade hint for hobby users, friendly errors for 402 / 429
- Pricing page: AI-model line per tier (Open-tier / Haiku / Sonnet / Opus),
Team €149 -> €199, Enterprise €499 -> €999, daily-preview limit per tier
- Privacy + Security: explicit subprocessor disclosure for Anthropic (US) /
Zhipu (CN) and which tier uses which
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Tailwind v4's `bg-[--color-X]` bracket-arbitrary syntax does not wrap the
value in var(), so it compiles to `background-color: --color-bg-elevated`
— an invalid color, which the browser falls back to transparent. The
mobile menu was the one element that depended solely on this utility for
its background, so it rendered with none.
Use an inline style with explicit var() and color-mix to match the nav
bar's frosted look (var(--color-bg) at 80% + backdrop-blur).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The dropdown was bg/95 + backdrop-blur — fragile across mobile browsers
where backdrop-filter is unreliable, leaving 5% transparency that read as
"no background". Switch to a solid elevated panel with a soft shadow and
an explicit z-index.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
On phones the dashboard top bar is tight with the nav icons + the primary
action crammed alongside. Move the action into a sticky bottom bar in the
thumb zone, leave the top bar to navigation. Hidden on the create-wizard
route since that page owns its own action.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sonnet still overran Cloudflare's edge timeout — the 504 fired at 90s but
the proxy had already cut the connection, so the browser saw a headerless
524 reported as a CORS error.
Measured against the live API: Haiku 4.5 generates the spec at ~200 tok/s,
so a full 8k-token spec completes in ~40s. With a hard 60s timeout and no
retries the route is guaranteed to answer well inside the proxy window.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The /v1/servers/preview route ran claude-opus-4-7 synchronously; full spec
generation routinely exceeded Cloudflare's ~100s proxy cap, so the browser
received a headerless 524 and reported it as a CORS failure.
- preview now uses claude-sonnet-4-6 with a 45s per-attempt timeout and one
retry — comfortably inside the proxy budget
- generateSpec maps an exhausted timeout to SpecTimeoutError; the route
returns a clean 504 (with CORS headers) instead of a stalled connection
- analyze step: live elapsed-seconds counter as freeze-proof, plus a
reduced-motion exception so the loading spinner keeps spinning (a status
indicator, which WCAG exempts from reduced-motion)
- textarea resize grip restyled to dark theme (light hatch on dark square)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- login: SMS step now has a 60-country dial-code <select> (CH default)
and a national-number input, combined into strict E.164 client-side
- marketing header: probe /v1/auth/me, show "Dashboard" when signed in
instead of the Sign in / Start building CTAs
- dashboard overview: drop the duplicate "+ New server" button, the
navbar one is the single source
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Hero h1 was a fixed text-[44px] — overflowed narrow phones. Now
text-[30px] sm:text-[40px] md:text-[56px].
- Hero grid children get min-w-0 so the code blocks' overflow-x-auto
actually constrains instead of widening the page.
- Marketing nav: the inline links were hidden below md with no fallback.
Added a hamburger MobileMenu; "Sign in" collapses into it on the
smallest screens.
- Section vertical padding is now responsive (py-14 sm:py-20).
- globals.css: overflow-x: clip on <html> as a safety net.
- docs: the 240px sidebar is hidden below lg, article gets min-w-0.
- dashboard header: nav labels collapse to icons on small screens.
Verified: next build passes (40/40 pages).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The domain was parked on GoDaddy Airo, which registered a Workbox
service worker. It keeps serving cached GoDaddy pages in browsers that
visited the parked domain. Serving a self-destruct sw.js makes those
browsers wipe the caches and unregister the worker on their next visit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- experimental.inlineCss: drop the render-blocking CSS request — the
Tailwind bundle is inlined into the HTML head (faster FCP/LCP on mobile).
- browserslist pinned to modern engines so Next/SWC stops emitting
polyfills for Baseline features (Array.at, Object.fromEntries, …).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
White on #6366f1 was 4.47:1 — just under the 4.5:1 minimum for small
text (Lighthouse a11y flag). Darkened the banner to #4f46e5 (6.3:1).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ported and adapted from the BuildMyDiscord SEO setup:
- lib/seo.ts — single source for site constants, the FAQ data (shared by
the rendered FAQ and the FAQPage schema so they never drift) and JSON-LD
builders.
- Rich root metadata: title template, keywords, Open Graph, Twitter card,
robots directives, canonical.
- JSON-LD: Organization + WebSite + SoftwareApplication sitewide, FAQPage
on the landing page. No AggregateRating — there are no real reviews yet.
- app/robots.ts — allow all, explicit allow-list for AI answer-engine
crawlers (GPTBot, ClaudeBot, PerplexityBot, …), disallow private routes.
- app/sitemap.ts — every public marketing + docs route.
- app/opengraph-image.tsx — monochrome on-brand 1200x630 share card.
- app/manifest.ts + public/llms.txt.
- Per-page metadata for pricing, changelog, security, privacy, terms,
docs, templates and status.
- opengraph-image + apple-icon pinned to the edge runtime — next/og
crashes during a Node-runtime prerender.
Verified: next build passes; /robots.txt, /sitemap.xml,
/manifest.webmanifest and /opengraph-image all generate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
bg-[--color-accent] does not resolve under Tailwind v4 — the banner bar
showed near-black. Set #6366f1 inline so the preview notice is clearly
visible regardless of theme wiring.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Clear notice that the service is not yet open for production use.
Temporary — remove SiteBanner once live.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lets Cloudflare run in Full mode (encrypted Cloudflare<->origin) instead
of Flexible (plaintext origin hop). Full (strict) is a later swap to a
Cloudflare Origin Certificate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The servers bind IPv4 (0.0.0.0) only. busybox wget resolves `localhost`
to ::1 first and does not fall back to IPv4, so the healthcheck failed
with "connection refused" and the container showed as unhealthy while
serving fine. Verified on the production api container.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Server recon (read-only SSH) showed the box already runs ~8 apps behind a
host-level nginx, with Gitea + an Actions runner. The host-networking
design collided with contentra on port 3001.
- docker-compose.prod.yml: bridge networking + per-app network, house
style; api/web/postgres/redis publish to 127.0.0.1 on verified-free
ports (4000/4001/5440/6390); only the generator keeps host networking
(no listening port, needs the host namespace for runner-port probing).
- Drop the Traefik config; the box uses a host nginx. Add a ready nginx
vhost in infra/nginx/buildmymcpserver.conf (listen 80, Cloudflare TLS).
- Add .gitea/workflows/deploy.yml mirroring the buildmydiscord pipeline.
- Narrow the generated-MCP port range to 4400-4900 (clear of screencraft
on 4321).
- .env.production.example + DEPLOY.md rewritten for buildmymcpserver.com
and the real topology.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The request said buildmymcp.com; the GoDaddy tab and the repo are named
buildmymcpserver.com. Added a top-of-file callout so the domain is resolved
before any DNS/nameserver change rather than baked in wrong.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The marketplace is the distribution channel — fork a working server or
publish your own — but it was absent from the landing page. Adds a
section between Examples and Pricing with a second conversion path into
/templates.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Multi-stage Dockerfiles for web/api/generator (pnpm workspace install,
tsx runtime — workspace packages are raw TS, same model as runner-template).
- docker-compose.prod.yml: postgres + redis + the three app services.
api/generator/web use host networking so the generator's host-port probe
is correct and every service shares one address space; api + generator
mount the Docker socket. Binds nothing on 80/443 — safe beside other apps.
- Optional Traefik reverse proxy in infra/traefik/ (heavily gated — only if
the box has no existing proxy).
- .env.production.example, .dockerignore, DEPLOY.md (Cloudflare zone, GoDaddy
nameserver switch, server deploy, Google Cloud Console OAuth app).
- api/generator `start` now runs via tsx; `node dist/index.js` could never
resolve the raw-TS workspace imports.
All three images verified building clean; the API container boots under tsx.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
/servers/new and /login/callback call useSearchParams() directly, which
bails the page out of static rendering and fails `next build` during
prerender. Split each into a thin Suspense wrapper + inner component.
Latent since `next dev` never prerenders — only surfaces in a prod build.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Server-side authorization-code flow: /v1/auth/google redirects to the
consent screen with a CSRF state cookie; /v1/auth/google/callback
exchanges the code, validates the ID token (iss/aud/exp/email_verified),
and mints a 30-day session via upsertOAuthLogin. /v1/auth/providers lets
the login UI hide the button until GOOGLE_OAUTH_ID/SECRET are set.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes structural weakness #4 from the audit (single global key, no rotation,
no KMS path). Customer secrets now use envelope encryption with a real
rotation story.
Model:
KEK — Key Encryption Key, 32 bytes from env (SECRETS_ENCRYPTION_KEY). Never
stored in the DB. Root of trust.
DEK — Data Encryption Key, 32 random bytes we generate, stored in the new
encryption_keys table *wrapped* (AES-256-GCM encrypted) with the KEK.
Secrets are encrypted with the DEK.
Schema:
- encryption_keys (version, wrappedDek, active, rotatedBy, createdAt, retiredAt)
- secrets.keyId — which DEK encrypted this row. NULL = legacy (KEK-direct,
pre-envelope); decryptSecret handles both and the first rotation migrates
legacy rows onto a DEK.
crypto.ts (full rewrite):
- ensureActiveKey() — boot-time, loads keys + creates v1 if none. Fail-closed:
index.ts process.exit(1) if it throws — the API will not serve if encryption
can't initialize.
- encryptSecret() — encrypts with the active DEK, returns { value, keyId }.
- decryptSecret(value, keyId) — DEK path or legacy KEK-direct path.
- rotateKeys() — mints a fresh DEK, re-encrypts EVERY secret under it inside a
single transaction (decrypt-old / encrypt-new per row), retires the old key,
activates the new one. A partial failure is recoverable because every row
carries its own keyId.
- encryptionStatus() — active version, key history, secret + legacy counts.
Admin:
- GET /v1/admin/encryption — status
- POST /v1/admin/encryption/rotate — triggers rotateKeys, audit-logged as
admin.encryption.rotate with { newVersion, reEncrypted }.
- /admin/encryption page — active-key/secret/legacy cards, Rotate button with
confirm, key-history table, plain-English how-it-works. Added to admin nav.
Verified end-to-end:
- boot → encryption_keys v1 active, '[crypto] envelope encryption ready'
- created a server with secret MY_API_KEY → stored ciphertext, keyId = v1
- POST rotate → { newVersion: 2, reEncrypted: 1 }; ciphertext changed, keyId
now v2, v1 retired, v2 active. The decrypt-then-reencrypt round-trip
succeeded (rotation throws otherwise) — the secret is provably recoverable.
- admin UI renders the status + history correctly.
Deferred, named honestly (not built this iteration):
- worker reads secrets from the DB instead of the BullMQ job-data plaintext
copy — would also remove plaintext secrets from Redis. Separate change with
its own risk surface on the iterate/fork flows.
- per-server secret-value rotation UI
- audit_log hash-chaining (tamper-evidence)
- rate limiting on auth endpoints
Sovereign-audit follow-up. The audit's finding pass missed this: every
Iterate (version > 1) ran allocatePort -> a NEW port and deployContainer -> a
NEW container, then pointed the DB row at it — and never stopped the old
container. The previous version kept running forever, holding a host port,
with the old secrets baked into its env, untracked (its containerId was
overwritten in the DB by deployContainer). Same bug class as API-SERVERS-001
but on the iterate path.
Fix: the worker captures the server's current containerId before the build
mutates the row, and after the new container is confirmed live + the DB
updated, it stops the old one. This also makes the 'rolling deploy' the UI
promises actually true — the old version stays up until the new one is live,
then is retired.
deploy.ts stopContainer now returns { ok, detail } (was void) so the worker
can log the outcome.
Verified: generator typecheck clean.
Full reasoning-based audit of all 10 zones. 11 findings, all confirmed real,
zero false positives. 5 fixed now, 6 deferred to a justified backlog.
API-SERVERS-001 (HIGH) — DELETE /v1/servers/:id orphaned the container
The route deleted the DB row but never stopped the Docker container — it
kept running forever on its host port, still serving traffic with the
user's secrets baked into its env. The takedown path got stopContainer in
an earlier commit; this sibling path was missed. DELETE now tears the
container down first. Verified: deleted 'gfgfg' — container 23e0c55c gone,
:4110 connection-refused after.
INFRA-001 (HIGH) — SECRETS_ENCRYPTION_KEY zero-default usable in production
The AES-256-GCM key defaults to 64 zeros and passes the min(64) check. A
prod deploy that forgot to set it booted silently with every secret
encrypted under a public key. config.ts now throws on boot when
NODE_ENV=production and the key is still the placeholder. Verified: prod
boot with the zero key is REFUSED.
API-SERVERS-002 (MEDIUM) — WS build stream had no authorization
GET /v1/builds/:id/stream streamed build logs with no auth, while its REST
twin checks orgId. Now authenticates from the session cookie and rejects
builds outside the caller's org. Verified: no cookie -> 'unauthorized';
cross-org build -> 'not_found'; own build -> streams (no regression).
OAUTH-001 (MEDIUM) — authorization code consumption was not atomic
The 'already used?' check and the 'mark used' write were separate
statements — two requests racing the same code could both mint tokens.
Now a conditional UPDATE ... WHERE consumed_at IS NULL RETURNING; the
loser of the race gets zero rows and invalid_grant.
OAUTH-002 (MEDIUM) — 'plain' PKCE accepted, contradicting AS metadata
The AS metadata advertises code_challenge_methods_supported: ['S256'] but
/oauth/authorize accepted 'plain'. Authorize is now z.literal('S256') and
pkceVerify dropped the plain branch. Verified: authorize with plain -> 400.
Deferred to backlog (documented in TEMPLATE_SECURITY_AUDIT.md is template-only;
this audit's findings are in the commit + certification):
GENERATOR-001 — secrets via docker -e (visible in docker inspect); needs
--env-file rework
RUNNER-001 — generated containers run as root; needs USER node + build
re-test
AUTH-001 — no rate limit on magic-link / oauth register; needs
@fastify/rate-limit
GENERATOR-002— allocatePort check/bind race; low, self-heals on rebuild
AUTH-002 — expired magic_links/sessions/oauth rows never purged; needs
a cron
FEATURES-001 — tool-call metering not wired (metrics always 0); Sprint 4
by plan
Bug: forking a template POSTed /v1/servers with slug = trySlug(template.title),
a fixed value. If the user had any server with that slug already (e.g. forking
the same template twice, or a name collision), the create returned 409
slug_taken — and the fork wizard skips Step 1, so there was no slug field to
fix it with. The user was stuck (8 repeated 409s in the report).
Fix:
- On fork setup, after the fork call, GET /v1/servers and auto-unique the
default slug: echo-demo-template -> echo-demo-template-2 -> -3 ... against the
user's existing slugs. Lookup failure is non-fatal (slug field is editable).
- Fork step 2 now renders editable Name + Slug fields in the fork banner
('must be unique in your workspace' hint) — the normal wizard has these in
Step 1, which the fork flow skips, so they belong here.
- slug_taken build error now reads 'The slug "x" is already used by one of
your servers — change the Slug field above' instead of the raw code.
Note: the SES lockdown-install.js and content.js 'query' errors in the report
are browser extensions, not the app.
Verified: forked echo-demo-template (whose base slug was already taken) — slug
auto-filled echo-demo-template-2, build succeeded, container live on :4111,
template fork_count incremented to 2.
The logged-in user can now reach the marketplace and filter to their own
templates.
Dashboard nav:
- Added 'Marketplace' item (Overview · Servers · Marketplace · Audit · Settings).
/templates page — login-aware:
- Detects session via /v1/auth/me. Logged-in users get a 'Dashboard' + '+ New
server' header instead of 'Home' + 'Start building'.
- New [All templates | My templates] scope toggle, shown only when logged in.
- 'My templates' loads GET /v1/templates/mine and shows EVERY status the user
owns (public / hidden / draft / takedown) with a colored status badge on each
card — so a template you unshared doesn't appear to have vanished.
- Sort tabs (trending/top/newest) hide in 'mine' scope — meaningless for a
handful of own templates. Category filter + search still apply (client-side).
- Takedown cards link to the source server's Publish tab instead of the detail
route (which 410s); everything else opens the detail page.
Backend:
- GET /v1/templates/mine (requireAuth) — all own templates, any status,
registered before /:slug so the static route always wins the match.
- GET /v1/templates/:slug — now does an optional session check: the OWNER can
view their own hidden/draft template (so a 'My templates' card click never
dead-ends in a 404). takedown stays 410 for everyone, owner included — that's
an admin decision, not the owner's to reverse.
Detail page:
- Fork CTA is gated on status === 'public'. For a non-public template the owner
sees an amber 'not forkable — re-share from the Publish tab' notice plus a
'Manage in server' link, instead of a Fork button that would fail silently.
Verified:
- GET /v1/templates/mine → marco's 1 template; 401 without auth
- Owner GET of a hidden template → 200 status:hidden; anon → 404
- Dashboard nav shows Marketplace (screenshot)
- /templates 'My templates' toggle → only own template, public badge, sort tabs
hidden (screenshot)
Goal: maximize template volume without a dark pattern and without leaking data.
Wizard Done-page Share panel:
- 'Share as template in the marketplace (recommended)' checkbox, default ON,
rendered inline in the build-success flow where every user lands.
- Honest copy — corrected a draft that claimed 'only abstracted code pattern is
shared'. That is false: the FULL generated code becomes publicly viewable on
the template detail page (by design, for pre-fork audit). The panel now says:
'Your secrets stay private ... but your generated code becomes publicly
viewable so others can audit it before forking. Unshare anytime.'
- When checked: inline minimal form — short description (prefilled from the
spec), category select, optional per-secret credential hints. One 'Publish to
marketplace' click. Not auto-published silently — that would be a consent dark
pattern; one visible deliberate click keeps it clean.
- Forked servers don't show the panel (re-publishing a fork is an edge case).
Owner unshare/reshare:
- GET /v1/servers/:id/template — owner lookup, drives the Publish tab UI.
- PATCH /v1/templates/:slug/visibility { shared } — owner-only toggle between
public and hidden. 403 for non-owners, 409 if an admin took it down (owner
cannot resurrect an admin takedown). Audit-logged as template.unshare /
template.reshare.
- Server-detail Publish tab now detects an existing template and shows the
shared status (public/hidden/takedown badge), fork count, a marketplace link
and an Unshare/Re-share button — instead of the publish form.
Why this is safe to default ON:
- Secrets are architecturally bound to mcp_servers, never copied into templates.
Publish reads tools_schema + generated_code only; the secrets table is never
touched. Data leak is structurally impossible, not policy-dependent.
- Publish re-scans the generated code for banned patterns AND hardcoded
credentials (sovereign-audit hardening) before it can reach the marketplace.
- The user sees a visible, pre-ticked checkbox and reads one honest sentence
before publishing. Privacy-conscious users untick; everyone else contributes
volume. Informed consent, GDPR-clean.
Verified end-to-end via API:
GET server/:id/template -> null (unpublished)
POST /v1/templates -> published, slug share-test-server
GET server/:id/template -> status public
PATCH visibility {shared:false} -> hidden, drops out of public list
PATCH visibility {shared:true} -> public again
UI: Publish tab renders the shared-status panel with View + Unshare (screenshot
confirmed).
Also: hero badge date set to 2026-05-20. Changed 'MCP spec 2025-11-25' to
'updated 2026-05-20' — claiming an MCP spec dated today would be factually wrong
(no such spec release exists); 'updated' is accurate and gives the requested
fresh date. The real spec date is still cited correctly in /docs.
P0 — three critical issues found by tracing every attack vector on the template
publish + fork + render path. All three fixed and verified with attack tests.
FIX A — Takedown actually stops malicious containers
PATCH /v1/admin/templates with status=takedown previously only updated
mcp_servers.status to 'paused' in the DB. The Docker container kept running
and serving traffic on its allocated port — takedown was cosmetic. Now the
endpoint enumerates every fork's container, calls 'docker rm -f' on each,
clears container_id/public_url/host_port in the DB, and returns the
stoppedContainers count. New apps/api/src/lib/docker.ts owns the stop logic.
Verified: takedown stopped container f5632962, port 4109 connection refused.
FIX B — Reject specEdit on fork
A hand-crafted POST /v1/servers with {templateId, previewId, specEdit} would
enter the spec-edit branch, merge edits into the cached spec, but the worker
reads the pre-built template code (separate cache key), ignoring the merged
spec entirely. User thinks they changed something; deployed container behaves
as the original. Now returns 400 spec_edit_forbidden_on_fork with an explainer
pointing to the Iterate flow.
FIX C — templateId validation via Redis fork-ref
templateId on POST /v1/servers was user-controlled and unvalidated:
fork_count of any template could be pumped, mcp_servers got garbage
template_id rows, takedown cascade would miss the bogus rows. Fork endpoint
now writes a Redis key fork-ref:<previewId> -> templateId (5min TTL).
Server-create requires the ref to exist AND match the submitted templateId.
Verified attack: fake templateId without fork-ref returns 410 fork_ref_expired.
DEFENSE-IN-DEPTH — Hardened static checks
Banned patterns (added):
Function\s*\(['"`] — Function('code')() form, no 'new' needed
\bimport\s*\( — dynamic import escapes bundle scope
\bsetTimeout\s*\(['"`] — setTimeout('code', ms) eval form
\bsetInterval\s*\(['"`]
\bfs\s*\.\s*(unlink|rmdir|rm)\b
\bprocess\s*\.\s*kill\b
you are now in (developer|jailbreak|dan) mode — extra jailbreak markers
Hardcoded-credential patterns (new — scanForLeakedSecrets):
sk-ant-(api|sid)… — Anthropic
sk-… — OpenAI
sk_(live|test)_… — Stripe
ghp_… — GitHub PAT
github_pat_… — GitHub fine-grained
xox[bpoasr]-… — Slack
AKIA[0-9A-Z]{16} — AWS
-----BEGIN…PRIVATE KEY----- — RSA / SSH / GPG
Triggered when a publisher pasted their key into the prompt and Claude
embedded it literally in the generated code. Publish-blocking.
Verified attack: smuggled 'Function("return 1")' into a build's
generated_code, attempted publish → 422 publish_blocked.
Slug regex tightened — fork + detail routes now require
^[a-z0-9][a-z0-9-]{0,63}$ (was loose min(1).max(64) — letting through
'../admin', long strings, mixed case).
UI warning — Publish-as-template form now shows an amber callout listing
what's scanned and explicitly stating egress allowlisting is roadmap, not
enforced today (was misleading: the field was collected, never enforced).
TEMPLATE_SECURITY_AUDIT.md added — documents all 20 audited vectors with
severity, status, and rationale for what's deferred.
UI polish
globals.css — select/input/textarea/button get color-scheme: dark + custom
chevron + option styling so Chrome's native popdown stops rendering as a
white OS-themed widget on dark pages. The /templates category dropdown was
the immediate trigger; same rule applies system-wide.
What this enables:
- A user builds an MCP server. If others would benefit, they click 'Publish as
template' on their server detail page. The spec + pre-rendered TypeScript
snapshot is preserved.
- Visitors browse /templates, filter by category, sort by trending/top/newest.
Each template card shows fork count + active deployment count as natural
manipulation-resistant popularity signal.
- /templates/[slug] shows the full plan: tool list with input schemas,
required-credential explanations (with 'how to get one' deep links), and a
collapsible code preview so users can audit before forking.
- Fork is one click → /servers/new?template=slug. The wizard skips Step 1 and
pre-fills Step 2 with the template's parsed spec. Forker only fills in their
own credentials. mcp_servers.template_id is recorded; template.fork_count is
bumped atomically. Each fork gets its own isolated container with its own
port, its own AES-256 secrets — the template author has zero visibility into
the fork's traffic or data.
- Admin /admin/templates moderation: verify quality templates (shows shield
badge in marketplace), hide low-effort ones, takedown anything malicious.
Takedowns cascade-pause every fork container — owners must re-deploy.
Why template+fork instead of shared-container:
- Shared containers would mean the publisher's quota + their secrets + their
logs are exposed to forkers. Bad ergonomics, bad security, bad ownership.
- Templates/forks decouple the spec (shared, vouched-for) from the runtime
(isolated per user). Network-effect moat without the trust collapse.
Why no 5-star voting in v1:
- Manipulation-anfällig, empty lists without adoption. We use fork count +
active deploys + verified badge. Trending algorithm:
score = (activeDeploys * 3 + forks) / sqrt(ageDays + 1)
Real signal, no brigading attack surface.
Backend:
- New schema: templates table (16 cols incl. tools_schema, generated_code,
required_secrets, allowedDomains, status enum, verified, fork_count).
- mcp_servers.template_id FK + idx for fork lookup.
- @bmm/types: SpecEdit unchanged, CreateServerInput accepts optional templateId.
- preview-cache.ts: new cachePrebuiltCode/loadPrebuiltCode for storing the
template's full rendered server.ts alongside the spec. Generator worker
detects this and skips the render step — uses the audited pre-built code
verbatim. Banned-pattern re-scan at publish time.
- routes/templates.ts: 5 public/auth routes + 2 admin routes. Banned-pattern
re-scan before publish. Slug auto-uniqued. forkCount atomic-increment via
SQL.
UI:
- /templates marketplace with trending/top/newest tabs, category filter, search.
Cards show forks + live count + author + verified badge.
- /templates/[slug] full detail with tools, credentials-with-hints, expandable
code preview, fork CTA, ownership + stats sidebar, 'forking is safe' explainer.
- /servers/new?template=slug — wizard auto-jumps to Step 2 with template spec
pre-filled, fork banner at top with link back to template.
- /servers/[id] new Publish tab with title, category, descriptions, per-secret
hint fields (description + howToGetUrl per UPPER_SNAKE_CASE key).
- /admin/templates moderation with verify/hide/takedown actions.
- Marketing nav now includes /templates.
Verified end-to-end:
- Published Echo Demo Template from marco@test.local's live server
- Marketplace lists it correctly with stats
- Detail page renders with all sections
- Fork CTA navigates to wizard with ?template= param
- Wizard skips Step 1, shows fork banner, pre-fills spec
- Build succeeds in ~10s (cached spec + prebuilt code path skips Claude AND
render), container live on :4109 with proper OAuth 401 → token → 200 flow
- DB: templates.fork_count=1, activeDeployments=1, mcp_servers.template_id
populated on the fork
- /admin/templates shows the new template with verify/hide/takedown controls
The indigo filled square didn't match the in-page Logo component. The nav-bar
logo is a monochrome outlined rounded square + M path, currentColor on the
dark page background.
Favicon now follows the same design: outlined rect + M, stroke adapts:
- light browser tabs → #0A0A0B (near-black)
- dark browser tabs → #FAFAFA
Apple-icon stays as the indigo filled tile — iOS home-screen icons need solid
backgrounds, monochrome outlines disappear there.
app/icon.svg — 32px vector with the brand 'M' (logo-matching) on #6366F1 indigo
rounded square. Sharp at every browser size.
app/apple-icon.tsx — 180x180 PNG rendered at request time via next/og
ImageResponse with the same design scaled up. Covers iOS home-screen + iPadOS.
Next 15 auto-discovers both via the file-based metadata convention and injects:
<link rel='icon' href='/icon.svg' type='image/svg+xml' sizes='any'>
<link rel='apple-touch-icon' href='/apple-icon' type='image/png' sizes='180x180'>
Verified: both URLs return 200, both link tags appear in the rendered HTML head,
brand matches the in-page Logo component.
The wizard's confirm step is no longer read-only. Users can refine what Claude
parsed before committing to a build.
Backend:
- @bmm/types adds SpecEdit (tools[name,description,inputSchema] + requiredSecrets);
CreateServerInput accepts an optional specEdit alongside previewId.
- Servers create endpoint: when specEdit is provided, loads cached spec from Redis,
index-merges the edits in (keeping LLM-generated implementations untouched),
re-validates via GeneratorSpec, re-runs the banned-pattern scan, overwrites the
Redis cache so the worker reads the user's version. Refuses with
preview_expired/tool_count_mismatch/banned_pattern on safety failures.
- New overwriteSpec() helper in preview-cache.
Frontend:
- Step 2 renders each tool as an editable card: name input, description textarea,
JSON schema textarea with parse-on-keystroke validation (inline error if invalid).
- Required secrets list is editable: keys via uppercase-snake-case input, +Add /
remove buttons, secret values kept in sync when keys are renamed.
- Reset-to-AI-suggestion button appears when edits are dirty.
- Pre-submit validation: schema must parse, secret keys must match UPPER_SNAKE_CASE,
required secret values must be provided.
- Warning copy: 'Renaming parameters may require an Iterate after build — the
existing impl references the original names.'
Verified end-to-end via browser smoke test: edited description + renamed tool
landed correctly in mcp_servers.tools_schema and in the live container at :4107.
Implementation field preserved from the original cached spec.
Sprint 3.5: close every dead link and replace the single-step wizard with the
spec-mandated 3-step flow.
Wizard:
- Step 1 collects prompt + name + slug, calls /v1/servers/preview.
- Step 2 renders parsed tools (name, description, input schema as copyable JSON)
+ a credential field per requiredSecret Claude actually identified. Self-contained
servers see 'No credentials needed' instead of generic Notion placeholders.
- Step 3 streams the live build over WebSocket and shows install snippets.
New dashboard pages:
- /settings — org, plan/usage, members table, API keys + billing stubs (Sprint 4),
encryption status. Reads /v1/me/org.
- /audit — filterable table over /v1/audit with action pills, resource refs, IP,
metadata JSON.
Docs site (/docs + 6 sub-pages):
- Sticky 240px sidebar, max-w-prose article column, shared DocsTitle/H2/Code primitives.
- Quickstart, MCP concepts, OAuth 2.1 flow (full walkthrough with curl), Authoring
tools, Self-hosting, API reference, FAQ.
Marketing pages:
- /changelog with tagged release timeline.
- /security with 8 pillars + disclosure.
- /privacy with GDPR-aware sections.
- /terms (10 clauses).
- /pricing full page (nav now points here instead of /#pricing anchor).
- /status with live 10s probes against /api/health and /login.
Footer 'system status' badge now links to /status.
All 20 routes 200 OK in smoke crawl. Typecheck clean across packages.
- POST /v1/servers/preview runs Claude synchronously, validates output, caches spec
in Redis under preview:<id> with 5min TTL, returns previewId+spec+detectedSecrets.
- POST /v1/servers accepts optional previewId; worker reuses the cached spec if
the entry is still present, otherwise regenerates fresh. Skips the second
Claude round-trip (~30s saved on the demoable path).
- audit() helper writes auth.login, auth.logout, server.create, server.iterate,
server.delete to audit_log with ip, metadata, resourceId.
- GET /v1/me/org returns organization + members list for the settings page.
- GET /v1/audit?limit=&action=&resourceType= returns scoped audit entries.
- Bump @modelcontextprotocol/sdk from 1.0.4 to 1.29.0 in runner-template
(1.0.4 has no McpServer or StreamableHTTPServerTransport — file not found at runtime).
- Bump zod to 3.25.76 across workspace to satisfy modern SDK peer dep.
- Split OAUTH_ISSUER (canonical, host-reachable) from CONTROL_PLANE_URL (container-reachable for JWKS).
Runner verifies iss against OAUTH_ISSUER; fetches JWKS from CONTROL_PLANE_URL.
Both API and runner now agree on http://localhost:4000/oauth as the issuer in dev.
- Move postgres host port 5432 to 5440, redis 6379 to 6390 to avoid collisions with
native installs on the dev machine.
- Move web from 3000 to 3001 (3000 occupied by Gitea on dev machine).
- Drop pino-pretty transport from API to avoid runtime require of an unbundled dep.
- Cast build_logs.level (varchar) to BuildEvent's literal union in WS replay path.
- Remove unused reqBase helper in oauth.ts.
