Tokenlab MCP Server

What hedging8563/tokenlab-mcp-server MCP server does

The hedging8563/tokenlab-mcp-server MCP server connects MCP clients to TokenLab's public catalog and broader developer API. Agents can list models, inspect model-specific operations and request formats, retrieve pricing, and compare several models before selecting an endpoint. The catalog profile is public-only, while the default core profile adds common inference, media, file, embedding, reranking, translation, and task operations. A full profile exposes the wider allowlisted API surface.

The server supports OpenAI-compatible Chat Completions, the OpenAI Responses format, Anthropic Messages, and Gemini GenerateContent. It also exposes image creation and editing, speech synthesis, transcription, audio translation, video, music, and 3D generation. File operations cover uploads, listing, retrieval, content retrieval, and deletion.

How it works

The package is generated from TokenLab's OpenAPI contract with an MCP-specific overlay. Tool definitions retain the relevant HTTP bindings, content types, authentication requirements, and endpoint-oriented argument structures. Calls are checked against the generated schema before an API request is sent. MCP responses include structured content while retaining serialized text for clients that do not consume structured results.

The server offers catalog, core, and full tool profiles through TOKENLAB_MCP_TOOL_PROFILE. Schema presentation can use portable, exact, or strict modes. Two prompts help agents choose a model from live information and construct a request without losing the selected native protocol shape. Resources provide the API overview, OpenAPI snapshot, and compact public contract.

Video, music, and 3D jobs are asynchronous. Image operations can be synchronous or asynchronous depending on the selected model. Creation responses provide normalized delivery information such as a task ID and polling URL; clients use get_task_status until the task reaches a terminal state. Larger binary results can be written to an artifact directory instead of being returned inline.

Setup and configuration

Install and run the package with:

npx -y @tokenlabai/mcp-server

A local installation can instead use npm install followed by npm start, or the project can be run from its Docker image. The documented Claude Desktop configuration starts the npm package with npx and can set TOKENLAB_API_BASE; the default API base is https://api.tokenlab.sh.

No key is needed for model discovery and pricing. Set TOKENLAB_API_KEY for credentialed inference, generation, asynchronous task, embedding, reranking, and translation calls. Other settings control the exposed profile, schema mode, request timeout, upload limit, inline response limit, and artifact output directory.

Tools and capabilities

The hedging8563/tokenlab-mcp-server MCP server includes tools for:

• Discovering models, model capabilities, pricing, and model comparisons.
• Sending Chat Completions, Responses, Anthropic Messages, and Gemini requests.
• Generating or editing images, speech, video, music, and 3D assets.
• Creating embeddings, multimodal embeddings, reranking documents, and translating text or audio.
• Uploading, listing, retrieving, and deleting files.
• Creating, polling, and cancelling supported asynchronous tasks.

Multipart operations accept local file paths. Chat image data URLs should use accurate image MIME types; recognized PNG, JPEG, WebP, and GIF data mislabeled as generic binary can be corrected at the MCP boundary.

Limitations and notes

Availability depends on the selected model's advertised request format and the provider route currently exposed by TokenLab. A model name or provider alone does not guarantee support for Responses or Gemini-native requests. Unknown fields are forwarded on a best-effort basis and remain subject to provider support.

Streaming-only and realtime operations are excluded because stdio MCP calls return one final result. Requests that accept streaming have it fixed internally to non-streaming. Cancellation currently applies only to queued Seedance video tasks for the documented model variants; processing, completed, failed, expired, and unsupported tasks are not cancelled. Image input and some multimodal features may require provider-specific enablement.