@stabgan/openrouter-mcp-multimodal

v5.1.3

Published

4 days ago

MCP server for OpenRouter — 21 tools: chat with 300+ LLMs, analyze & generate images/audio/video, TTS/STT, embeddings, reranking, structured decisions, async completions, response caching

What is this?

OpenRouter MCP Multimodal is a production-grade Model Context Protocol (MCP) server — listed on the official MCP Registry as io.github.stabgan/openrouter-multimodal. It connects AI coding agents (Cursor, Claude Desktop, VS Code, Windsurf, Cline, and others) to OpenRouter's unified LLM API over stdio.

Unlike text-only MCP servers, one install covers the full multimodal surface:

| Capability | Tools | Highlights | | :------------ | :-------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Chat | chat_completion, start_chat_completion, get_chat_completion_status | 300+ models, :nitro / :floor / :free / :online / :exacto suffixes, provider routing, web search, response caching, reasoning tokens, reasoning effort control, structured output (JSON Schema), async jobs, multi-model fusion, subagent delegation, response healing | | Vision | analyze_image, generate_image, generate_image_dedicated | OCR, captioning, VQA, image generation with reference inputs, dedicated Image API with resolution/quality/format control | | Audio | analyze_audio, generate_audio, text_to_speech, speech_to_text | Transcription, speech/music generation, dedicated TTS (Deepgram, ElevenLabs, OpenAI; free Deepgram default, voices, mp3/pcm), dedicated STT (Whisper, GPT-4o Transcribe, Voxtral, ElevenLabs Scribe) | | Video | analyze_video, generate_video, generate_video_from_image, get_video_status | Clip understanding, Veo 3.1 / Seedance 2.5 / Kling v3.0 / Wan 2.7 / Grok Imagine Video generation with progress notifications | | Embedding | embed_text | 37+ embedding models (text-embedding-3-small/large, Nomic, Jina, Voyage, etc.) | | Decisions | decide | Structured yes/no, choice, or score decisions with calibrated probabilities (System One models) | | Catalog | search_models, get_model_info, validate_model, rerank_documents, health_check | Model discovery, validation, reranking, ops health |

Production hardening: input/output path sandboxes (including analyze_* local files as of v4.5.2), SSRF guards, structured errors with _meta.code, MCP 2025-11-25 structured outputs + tool icons, async video progress notifications, and 3929 automated tests (3905 unit/mock + 7 regression + 17 live integration).

Quick start

1. Get an API key (free tier works) → openrouter.ai/keys

2. Run the server

export OPENROUTER_API_KEY=sk-or-v1-...
npx -y @stabgan/openrouter-mcp-multimodal

3. Add to your MCP client — copy one JSON block from Install into your client config:

| Client | Config location | | :----------------- | :-------------------------------------------------------------------------------------------------------------------------------- | | Cursor | Project: .cursor/mcp.json · User: Cursor Settings → MCP | | Claude Desktop | macOS: ~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json | | VS Code | .vscode/mcp.json (workspace) or User Settings → MCP | | Windsurf | Windsurf Settings → MCP (same mcpServers JSON shape as Cursor) |

Use the mcpServers object from Manual config below.

No credits required to start. Free models such as google/gemma-4-26b-a4b-it:free work for chat and vision. Video/audio generation typically needs credits.

Install

MCP servers are distributed through several packaging models. This server is implemented in Node.js/TypeScript; the table below maps each ecosystem method to how you run it here.

uvx vs npx: In the MCP ecosystem, npx runs npm (Node) packages and uvx runs PyPI (Python) packages. Because this server is Node-based, uvx uses a thin Python launcher that execs npx -y @stabgan/openrouter-mcp-multimodal — you still need Node installed.

One-click

Paste your OPENROUTER_API_KEY when prompted — deeplinks use placeholders so secrets never appear in URLs.

Manual config

export OPENROUTER_API_KEY=sk-or-v1-...
npx -y @stabgan/openrouter-mcp-multimodal

{
  "mcpServers": {
    "openrouter": {
      "command": "npx",
      "args": ["-y", "@stabgan/openrouter-mcp-multimodal"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-v1-..."
      }
    }
  }
}

Pin a release: "args": ["-y", "@stabgan/[email protected]"]

Install uv (includes uvx), ensure Node.js 22+ is also on your PATH, then:

export OPENROUTER_API_KEY=sk-or-v1-...
uvx mcp-server-openrouter-multimodal
# pin npm version: OPENROUTER_MCP_NPM_VERSION=5.1.3 uvx mcp-server-openrouter-multimodal

{
  "mcpServers": {
    "openrouter": {
      "command": "uvx",
      "args": ["mcp-server-openrouter-multimodal"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-v1-..."
      }
    }
  }
}

pipx equivalent: pipx run mcp-server-openrouter-multimodal

Optional: OPENROUTER_MCP_NPM_VERSION=5.1.3 pins the underlying npm package.

npm install -g @stabgan/openrouter-mcp-multimodal

{
  "mcpServers": {
    "openrouter": {
      "command": "openrouter-multimodal",
      "env": { "OPENROUTER_API_KEY": "sk-or-v1-..." }
    }
  }
}

git clone https://github.com/stabgan/openrouter-mcp-multimodal.git
cd openrouter-mcp-multimodal
npm ci && npm run build

{
  "mcpServers": {
    "openrouter": {
      "command": "node",
      "args": ["/absolute/path/to/openrouter-mcp-multimodal/dist/index.js"],
      "env": { "OPENROUTER_API_KEY": "sk-or-v1-..." }
    }
  }
}

docker run --rm -i -e OPENROUTER_API_KEY=sk-or-v1-... stabgan/openrouter-mcp-multimodal:latest

{
  "mcpServers": {
    "openrouter": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-e",
        "OPENROUTER_API_KEY=sk-or-v1-...",
        "stabgan/openrouter-mcp-multimodal:latest"
      ]
    }
  }
}

Use -i (interactive stdio). Avoid -t (TTY corrupts MCP framing on some hosts).

docker run --rm -i -e OPENROUTER_API_KEY=sk-or-v1-... \
  ghcr.io/stabgan/openrouter-mcp-multimodal:5.1.3

{
  "mcpServers": {
    "openrouter": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-e",
        "OPENROUTER_API_KEY=sk-or-v1-...",
        "ghcr.io/stabgan/openrouter-mcp-multimodal:5.1.3"
      ]
    }
  }
}

Interactive install (writes config for your client):

npx -y @smithery/cli install @stabgan/openrouter-mcp-multimodal --client claude
# or: --client cursor | vscode | windsurf | ...

Listing: smithery.ai/server/@stabgan/openrouter-mcp-multimodal

Official name: io.github.stabgan/openrouter-multimodal

Registry: registry.modelcontextprotocol.io
npm package: @stabgan/openrouter-mcp-multimodal
OCI image: docker.io/stabgan/openrouter-mcp-multimodal

Clients that support registry-driven install will offer npm or Docker; otherwise use the JSON blocks above.

claude mcp add openrouter -- npx -y @stabgan/openrouter-mcp-multimodal
# project scope:
claude mcp add --scope project openrouter -- npx -y @stabgan/openrouter-mcp-multimodal

Set OPENROUTER_API_KEY in your shell or client env before starting Claude Code.

Debug tools/list and tool calls against a live OpenRouter key:

export OPENROUTER_API_KEY=sk-or-v1-...
npx -y @modelcontextprotocol/inspector npx -y @stabgan/openrouter-mcp-multimodal

When Claude Desktop or Cursor cannot find npx (GUI apps often miss shell PATH), wrap with cmd:

{
  "mcpServers": {
    "openrouter": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "@stabgan/openrouter-mcp-multimodal"],
      "env": { "OPENROUTER_API_KEY": "sk-or-v1-..." }
    }
  }
}

If still failing, use the full path from where npx as the command.

Why this server?

| Capability | This server | Typical MCP LLM servers | | :----------------------------------- | :---------: | :---------------------: | | Text chat (300+ models) | ✅ | ✅ | | Image analysis + generation | ✅ | partial | | Audio analysis + TTS + STT | ✅ | ❌ | | Video analysis + generation | ✅ | ❌ | | Vector embeddings (37+ models) | ✅ | ❌ | | Structured decisions (System One) | ✅ | ❌ | | Model search / validate / rerank | ✅ | ❌ | | Path sandbox + SSRF protection | ✅ | rare | | MCP 2025 structured outputs | ✅ | rare | | Async video + progress notifications | ✅ | ❌ |

Tools

21 MCP tools. Each description includes Use when, Good/Bad examples, Fails when, and Works with so agents pick the right tool and recover from errors.

Every error carries _meta.code from a closed taxonomy (13 codes). Switch on these for programmatic handling:

Errors include _meta.suggestions with agent-oriented next steps when available.

Binary tool results (v4.7.0+)

Generate tools (generate_image, generate_image_dedicated, generate_audio, text_to_speech, generate_video, generate_video_from_image, get_video_status) return image, audio, or video bytes. As of 4.7.0 the behavior is explicit:

| save_path | Tool result | | :---------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Set | Text pointer only — e.g. Image saved to: out.png (… bytes, image/png) plus _meta.save_path. No inline base64 (avoids duplicating large payloads in the MCP channel). | | Unset, under byte ceiling | Inline media block and summary text (images/audio use MCP image / audio types; video uses MCP resource blocks). | | Unset, over ceiling | Text only with a hint to pass save_path. |

Default inline ceilings (override per kind or globally):

If you previously relied on both a saved file and inline media in the same tool result, read the file from _meta.save_path (or omit save_path to get inline media when under the ceiling).

Examples

Chat (free model)

{
  "tool": "chat_completion",
  "arguments": {
    "model": "google/gemma-4-26b-a4b-it:free",
    "messages": [{ "role": "user", "content": "Summarize MCP in one sentence." }]
  }
}

Chat with structured output (JSON Schema)

{
  "tool": "chat_completion",
  "arguments": {
    "model": "openai/gpt-4o",
    "messages": [{ "role": "user", "content": "Extract the name and age from: 'Alice is 30 years old.'" }],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "person",
        "schema": {
          "type": "object",
          "properties": {
            "name": { "type": "string" },
            "age": { "type": "integer" }
          },
          "required": ["name", "age"]
        }
      }
    }
  }
}

For simpler JSON without a schema, use "response_format": { "type": "json_object" }. Add "response_healing": true to auto-fix malformed JSON.

Analyze an image

{
  "tool": "analyze_image",
  "arguments": {
    "image_path": "diagram.png",
    "question": "List every label in this diagram."
  }
}

Use image_path and question — not image / prompt.

Search models (vision + free)

{
  "tool": "search_models",
  "arguments": {
    "query": "gemma",
    "capabilities": { "vision": true },
    "limit": 10,
    "offset": 0
  }
}

Generate video (async)

{
  "tool": "generate_video",
  "arguments": {
    "model": "google/veo-3.1",
    "prompt": "Ocean waves at sunrise, cinematic drone shot",
    "duration": 4,
    "save_path": "river.mp4"
  }
}

If the job is still running when max_wait_ms elapses, the response succeeds with _meta.code: JOB_STILL_RUNNING and a video_id — call get_video_status to resume. This is not an error.

With save_path set (as above), the result is a text pointer to the saved file once complete — not inline video. See Binary tool results.

Embed text (vector embeddings)

{
  "tool": "embed_text",
  "arguments": {
    "texts": ["What is MCP?", "Model Context Protocol enables tool use for LLMs."],
    "model": "openai/text-embedding-3-small",
    "dimensions": 1024
  }
}

Returns embedding arrays for RAG, semantic search, or clustering. 37+ models supported — pass a single string or an array of up to 2048 texts.

Structured decision (System One)

{
  "tool": "decide",
  "arguments": {
    "questions": {
      "is_bug": {
        "type": "noul",
        "instructions": "Is this a bug report or a feature request?",
        "criteria": {
          "true": "Describes broken or incorrect behavior",
          "false": "Requests new functionality"
        }
      }
    },
    "state": {
      "ticket": "The login button does not respond when clicked on Safari."
    }
  }
}

Returns calibrated probabilities instead of generated text. Three question types: noul (yes/no), choice (pick one category), score (ordinal rubric). Models: typesafe/jev-1.13 (default), upstage/solar-decide, microsoft/decision-1.

More examples: docs/plans/tool-description-improvement.md

Security

Input path sandbox — local paths on analyze_* and reference images must stay inside OPENROUTER_INPUT_DIR (falls back to OPENROUTER_OUTPUT_DIR, then cwd)
Output path sandbox — save_path must stay inside OPENROUTER_OUTPUT_DIR
Async job reads — get_chat_completion_status resolves disk paths only under OPENROUTER_OUTPUT_DIR/openrouter-jobs/ (4.7.0+)
SSRF protection — private/reserved IPs blocked on URL fetches
Untrusted content — analyze outputs tagged _meta.content_is_untrusted: true
Prototype-pollution defense — all API output boundaries (model records, embeddings, decisions, rerank results, log context) recursively strip __proto__, constructor, and prototype keys before forwarding to MCP clients
Secret redaction — the stderr JSON logger redacts API keys and sensitive values; stdout is reserved for MCP protocol traffic

Override sandboxes only with OPENROUTER_ALLOW_UNSAFE_PATHS=1 (discouraged).

Report vulnerabilities: SECURITY.md (private disclosure — do not file public issues for exploits).

Configuration

See .env.example for the full list (provider routing, fetch limits, caching, video polling, async jobs, integration-test overrides).

Development

git clone https://github.com/stabgan/openrouter-mcp-multimodal.git
cd openrouter-mcp-multimodal
npm install
cp .env.example .env   # add OPENROUTER_API_KEY
npm run build

Testing

Free models for CI / zero-credit accounts: integration tests default to google/gemma-4-26b-a4b-it:free (override with OPENROUTER_INTEGRATION_MODEL). GitHub Actions requires the OPENROUTER_API_KEY repository secret.

Mock tests live under src/__tests__/mock/ and cover handlers, path sandboxes, SSRF blocks, model-cache pagination, tool descriptions, and structured outputs — 396 additional cases beyond the core suite.

npm run lint
npm run format:check
npm run version:check   # package.json vs src/version.ts, server.json, pyproject.toml

Releasing

Published artifacts (npm, PyPI/uvx, Docker, GHCR) all ship from the same semver on a git tag (vX.Y.Z). Pushing to main runs tests but does not publish to npm or PyPI.

Normal flow: merge conventional commits to main → Release Please opens a Release PR → merge it → tag is created → CI publishes everywhere.

Manual flow: bump all version files → npm run version:check → npm run ci + smoke tests → commit → git tag vX.Y.Z → git push origin vX.Y.Z.

Full checklist, file list, CI secrets, and agent instructions:

docs/RELEASING.md — maintainer release guide
AGENTS.md — quick reference for AI agents

Troubleshooting

| Symptom | Likely cause | Fix | | :---------------------------------------------------------- | :--------------------------------------- | :-------------------------------------------------------------------------------------------------------------- | | Server exits immediately / OPENROUTER_API_KEY is required | Missing or empty API key | Set OPENROUTER_API_KEY in client env or shell — get one at openrouter.ai/keys | | _meta.code: INVALID_CREDENTIALS or HTTP 401 | Bad or revoked key | Regenerate at openrouter.ai/keys; restart the MCP client | | _meta.code: MODEL_NOT_FOUND | Typo or retired model ID | Run search_models or validate_model; check openrouter.ai/models | | HTTP 402 / insufficient credits | Paid model or generation on zero balance | Add credits at openrouter.ai/credits or use a :free model | | _meta.code: UPSTREAM_HTTP with 429 | Rate limit | Wait for _meta.retry_after_seconds if present; reduce concurrency | | _meta.code: UNSAFE_PATH | Local path outside sandbox | Put files under OPENROUTER_INPUT_DIR or set OPENROUTER_OUTPUT_DIR wider; see Security | | npx not found (Windows GUI apps) | GUI PATH differs from terminal | Use the Windows npx cmd /c wrapper | | No inline image/audio after upgrade | v4.7.0 with save_path set | Expected — result is text + _meta.save_path only; omit save_path or read the saved file | | MCP client shows stale tool list | Client cache | Restart MCP / reload window after upgrading the package pin |

Structured errors include _meta.suggestions with agent-oriented next steps when available.

FAQ

Do I need paid OpenRouter credits?

No, to get started. Free models work for chat and vision. Audio/video generation usually requires credits; analysis may return 402 on some models — the server surfaces that as a structured error.

Which MCP clients are supported?

Any MCP-compatible client over stdio: Cursor, Claude Desktop, VS Code Copilot, Windsurf, Cline, Kiro, and custom agents.

How is this different from calling OpenRouter directly?

This server adds MCP tool schemas, security sandboxes, error taxonomy, model caching, async video polling with progress notifications, and agent-oriented tool descriptions — so LLMs invoke the right capability without custom HTTP glue.

Where is the security advisory for path traversal?

Fixed in 4.5.2+ — see GHSA-3q7p-736f-x44v, SECURITY.md, and docs/solutions/security-issues/.

Compatibility

Works with any MCP client. Protocol: MCP 2025-11-25 (backwards-compatible with 2025-06-18 and earlier). Node ≥ 22 (Docker image uses Node 24).

License

Apache 2.0 — see LICENSE.

Contributing

Issues and PRs welcome. For large changes, open an issue first.

Before submitting: run npm run ci. Use Conventional Commits (fix:, feat:, etc.) so Release Please can cut the next release. See docs/RELEASING.md if you need to ship a version.