openrouter-mcp-multimodal

Purpose

OpenRouter MCP Multimodal is a production-grade multimodal MCP server used to connect AI coding agents (Cursor, Claude Desktop, VS Code, Windsurf, Cline, Kiro, etc.) to OpenRouter's unified LLM API, communicating via stdio. Unlike text-only MCP servers, a single installation covers the full chain of text, vision, audio, and video, and can both analyze and generate.

Main capabilities

• Text chat: can call 300+ OpenRouter models; the materials mention Claude, Gemini, GPT, etc.; supports structured output (JSON Schema), can also use response_format's json_object, and supports response_healing to automatically repair malformed JSON.
• Multimodal analysis: supports image, audio, and video analysis (the analyze_* series; local file analysis has been restricted by the input path sandbox since v4.5.2).
• Multimodal generation: image generation (generate_image, generate_image_dedicated), audio generation (generate_audio), text-to-speech (text_to_speech), video generation (generate_video, generate_video_from_image), and video status queries (get_video_status); the materials mention that generation models include Veo 3.1, Sora, Seedance, Wan. Video generation is asynchronous; if it is still running after max_wait_ms, it returns successfully with _meta.code: JOB_STILL_RUNNING and video_id, and get_video_status can be called to continue polling (this is not an error).
• Model search: can search models (for example, filtering for models that support vision or are free).
• Text embeddings: returns an embedding array for RAG, semantic search, or clustering; the materials state it supports 37+ models, and a single string or an array of up to 2048 texts can be passed.
• Structured decisions (System One): returns calibrated probabilities instead of generated text; supports three question types: noul (yes/no), choice (choose one category), score (ordered rating); models include typesafe/jev-1.13 (default), upstage/solar-decide, microsoft/decision-1.
• A total of 21 MCP tools, each tool description includes Use when, Good/Bad examples, Fails when, Works with, making it easier for agents to choose the correct tool and recover from errors.

How to use

1. Get an API key (free quota available): openrouter.ai/keys.
2. Run the server.
3. Add the JSON block to the MCP client configuration (use the mcpServers object in Manual config).

Installation/run methods listed in the materials:

• npx (Node package): @stabgan/openrouter-mcp-multimodal, can use "args": ["-y", "@stabgan/openrouter-mcp-multimodal@5.1.0"] to pin the version.
• uvx: because this server is based on Node, uvx uses a lightweight Python launcher to execute npx -y @stabgan/openrouter-mcp-multimodal, so Node still needs to be installed.
• pipx equivalent: pipx run mcp-server-openrouter-multimodal; OPENROUTER_MCP_NPM_VERSION=5.1.0 can be used to pin the underlying npm package version.
• Docker: OCI image docker.io/stabgan/openrouter-mcp-multimodal.
• Interactive installation (automatically writes the client configuration) and Smithery listing: smithery.ai/server/@stabgan/openrouter-mcp-multimodal.
• Official name: io.github.stabgan/openrouter-multimodal; Registry: registry.modelcontextprotocol.io.

Need to shell or the client environment sets OPENROUTER_API_KEY. When using uvx, you must first install uv and ensure Node.js 22+ is in PATH. When using npx, use -i (interactive stdio) and avoid -t (TTY breaks MCP framing on some hosts). If the GUI application cannot find npx, you can use a cmd wrapper or use the full path to npx.

Security and Limits

• Input path sandbox: local paths for analyze_* and reference images must be located in OPENROUTER_INPUT_DIR (falling back to OPENROUTER_OUTPUT_DIR, then falling back to cwd).
• Output path sandbox: save_path must be located in OPENROUTER_OUTPUT_DIR.
• Asynchronous task reading: get_chat_completion_status only resolves disk paths under OPENROUTER_OUTPUT_DIR/openrouter-jobs/ (4.7.0+).
• SSRF protection: private/reserved IPs are blocked during URL fetching.
• Untrusted content: analysis output is marked with _meta.content_is_untrusted: true.
• Prototype pollution defense: all API output boundaries recursively remove proto, constructor, and prototype keys.
• Secret redaction: stderr JSON logs redact API keys and sensitive values; stdout is reserved for MCP protocol traffic.
• The sandbox can only be overridden with OPENROUTER_ALLOW_UNSAFE_PATHS=1 (not recommended).
• Binary tool results (v4.7.0+): generation-type tools return image, audio, or video bytes, with a default inline limit (which can be overridden by type or globally); if save_path is set, the result is a text pointer to the saved file rather than inline media.
• Errors all carry _meta.code (13 closed categories), and where possible carry _meta.suggestions.
• Free models can be used for chat and vision; audio/video generation usually requires credits; some model analyses may return 402, which the server presents as a structured error.
• Integration tests use google/gemma-4-26b-a4b-it:free by default(can be overridden with OPENROUTER_INTEGRATION_MODEL); GitHub Actions requires the OPENROUTER_API_KEY repository secret.
• The tool list and calls can be debugged against a live OpenRouter key (tools/list).
• The protocol is MCP 2025-11-25 (backward compatible with 2025-06-18 and earlier); Node ≥ 22 (the Docker image uses Node 24); license Apache 2.0.

Applicable Scenarios

Suitable for developers and agent workflows that want to uniformly invoke multimodal capabilities through one MCP server in AI programming clients, for example in Cursor, Claude Desktop, and VS Code for multi-model conversations, analyzing local images/audio/video, generating images/speech/music/video, and doing text embeddings and structured decision-making.

Missing Information

The materials do not provide information such as specific pricing, the complete parameter list for each tool, the full details of all supported models, and the user counts for each platform.