🍳 ClaudioKitchen

Remote MCP server (Streamable HTTP) that gives Claude access to OpenRouter's image, video, TTS, transcription, embeddings, and rerank APIs. It adds OIDC auth (Pocket ID or any OIDC provider) with an email allowlist, plus file management and per-request cost tracking.

Why an OIDC proxy

Claude's remote MCP connectors require OAuth with Dynamic Client Registration (DCR). Pocket ID, like most OIDC providers, doesn't support DCR. FastMCP's OIDCProxy bridges the gap: Claude registers dynamically against the MCP server, and the server proxies the real login to your OIDC provider. Once you're logged in, ALLOWED_EMAILS is enforced on every tool call.

Related MCP server: OpenAI Agents MCP Server

Setup

1. Pocket ID: create a new OIDC client

   • Callback URL: https://<your-mcp-domain>/auth/callback

   • Copy the client id and secret

2. Copy .env.example to .env and fill everything in. BASE_URL must be the public HTTPS URL, with a reverse proxy in front (Traefik, Caddy, Pangolin, etc.).

3. docker compose up -d --build

4. In claude.ai: Settings → Connectors → Add custom connector, URL: https://<your-mcp-domain>/mcp. Claude redirects you to Pocket ID, you log in, and you're done.

Docker image

A GitHub Actions workflow (.github/workflows/docker-publish.yml) builds and pushes an image to the GitHub Container Registry on every push to main and on v* tags. To run the prebuilt image instead of building locally, point compose.yml at it:

services:
  openrouter-mcp:
    image: ghcr.io/pianonic/claudiokitchen:latest
    # drop the `build: .` line

Tags published: latest (default branch), the short commit SHA, and semver tags like 1.2 / 1.2.3 when you push a v1.2.3 tag. The package is private until you set it public in the repo's package settings.

Tools

Every model param accepts any OpenRouter model id. Discover them with list_models / list_video_models. Defaults are overridable via env (DEFAULT_IMAGE_MODEL, DEFAULT_TTS_MODEL, DEFAULT_STT_MODEL). Image-only models (Flux, Sourceful, Recraft, etc.) are handled automatically — if a model rejects the text modality, generate_image retries image-only. That includes Recraft's vector models (e.g. recraft/recraft-v4.1-pro-vector), which return a real SVG saved as .svg.

Per-request cost: every generating tool reports what the request cost in USD. generate_image and edit_image append a 💲 Request cost line; the rest return a cost_usd field. (OpenRouter credits are USD.)

Inline image display: generate_image and edit_image also ship an MCP Apps UI resource (ui://claudiokitchen/image.html). On hosts that support MCP Apps (claude.ai), the result renders inline in the chat instead of only as a download link. This works around a claude.ai limitation where a tool's image content is otherwise shown only inside the collapsed tool call. Hosts without MCP Apps support ignore it and fall back to the image block plus the download link, so nothing is lost.

Video runs in two modes:

• generate_video(..., wait=False) (default): submits the job and returns a job id. Poll it later with check_video.

• generate_video(..., wait=True, wait_timeout=600): polls internally until the clip is done, emitting progress notifications, then returns the finished result with the downloaded video URL. Good for short clips. If it exceeds wait_timeout it returns the job id so you can continue with check_video.

Note: a remote MCP server can only return data as the result of an active tool call, or as progress notifications during one. It cannot push an unsolicited message into a Claude chat after the turn ends. Use wait=True to have Claude report the finished video automatically.

Generated images, audio, and video are saved to FILES_DIR and served at /files/<uuid>.<ext>?token=<FILES_TOKEN>. Completed videos are pulled from OpenRouter's unsigned_urls onto this server, since those URLs expire.

Uploading your own files (images, PDFs, audio)

Any file type works. The extension comes from the filename you provide, or is sniffed from the bytes and content-type (falling back to .bin). Downloads are always served as attachments with nosniff. There are three ways to upload.

A) Claude with a sandbox (recommended, no base64): Claude calls create_upload_url to mint a short-lived ticket, then PUTs the raw file bytes directly:

curl -T /path/to/report.pdf "<upload_url>&name=report.pdf"

The PUT returns {"url": "https://.../files/<id>.pdf?token=..."}. Claude passes that url to the user, or to edit_image / describe_image / transcribe_audio / generate_video. Add &name=<filename> to keep the right extension. Tickets are time-limited (expires_in, default 900 s) and use-limited (max_uses, default 5).

B) Browser uploader (for a file pasted or attached in chat): open <BASE_URL>/upload?token=<FILES_TOKEN>, drag-drop any file, and copy the returned URL into chat.

C) upload_file tool: for a file Claude already has as base64, a data: URL, or an http URL. Pass filename to preserve the extension. Base64 decoding is tolerant, and the type is auto-detected (images, PDF, MP4, WAV, MP3, OGG, FLAC, ZIP, and more).

Cost tracking and budget

Every generating tool records its per-request USD cost to an append-only ledger (usage.jsonl, stored next to FILES_DIR on the volume; relocate it with USAGE_LOG). Ask for usage_summary to see today / this-month / all-time spend with a per-tool breakdown. Set MONTHLY_BUDGET_USD to a hard cap. Once this calendar month's recorded spend reaches it, generating tools refuse to run until next month. Video jobs are recorded once (deduped by job id) even if you poll check_video repeatedly.

Managing generated files

Files served under /files accumulate on the volume. Use list_files (name, size, modified time, URL, plus totals), delete_file (by filename or download URL), and cleanup_files (bulk-prune by older_than_days and/or keep_newest; dry_run=True previews first).

File security

The /files/ route requires ?token=<FILES_TOKEN>. Set FILES_TOKEN in .env (a stable secret) so download links keep working across restarts. If it's unset, a random one is generated per run and logged. Returned download URLs already include the token.

Hardening applied in this server:

• Downloads are served with X-Content-Type-Options: nosniff and Content-Disposition: attachment, so a stored .svg or .html can't execute script in the server's origin.

• Outbound fetches (upload_file from a URL, edit_image, describe_image, transcribe_audio) are restricted to public http(s) hosts. Loopback, link-local (including 169.254.169.254), and private ranges are blocked (SSRF guard), redirects are disabled, and a size cap applies (FETCH_MAX_BYTES, default 100 MB). Uploads are capped by UPLOAD_MAX_BYTES.

Known limitation: FILES_TOKEN is a single global secret carried in the URL query string. Don't pass /files URLs to third-party APIs, and configure your reverse proxy to strip query strings from access logs. Rotating it invalidates all existing download links.

Notes

• Video generation costs credits the moment you submit. Check pricing and capabilities with list_video_models first.

• TTS uses streamed pcm16 (OpenRouter requires stream:true for audio output), wrapped into a 24 kHz mono WAV.

• For local testing without HTTPS, tunnel with cloudflared or ngrok and set that URL as BASE_URL.

• A /health endpoint (unauthenticated) is available for container and reverse-proxy health checks. The Docker image wires it into a HEALTHCHECK.

License

Apache-2.0 © pianonic