Litellm Mcp

LiteLLM MCP Server

An MCP server for LiteLLM operations. At startup it reads the LiteLLM proxy OpenAPI document and registers one FastMCP tool per LiteLLM endpoint.

Features

• Full LiteLLM API coverage: Registers every operation in GET /openapi.json.
• Version-aware tools: Tool names come from LiteLLM operationId values, so

the exposed tools follow the running LiteLLM version.

• Discovery helpers: Includes tools to list/search registered LiteLLM tools

and inspect individual endpoint mappings.

• Environment-based config: Reads the LiteLLM URL and master key from env vars.

Prerequisites

• Python 3.10+
• uv package manager recommended
• A reachable LiteLLM proxy
• A LiteLLM master key or admin key

Installation

uv sync

Configuration

The server reads LiteLLM connection settings from environment variables. It fetches ${LITELLM_URL}/openapi.json during startup.

• LITELLM_URL: LiteLLM proxy base URL, for example http://localhost:4000.
• LITELLM_MASTER_KEY: LiteLLM master key or admin key.
• LITELLM_TIMEOUT_SECONDS: Optional request timeout in seconds. Defaults to 30.
• MCP_TRANSPORT: MCP transport, either stdio (default) or http.
• MCP_HOST: HTTP bind address. Defaults to 0.0.0.0.
• MCP_PORT: HTTP bind port. Defaults to 8000.

Usage

Running the Server

export LITELLM_URL=http://localhost:4000
export LITELLM_MASTER_KEY=sk-...
uv run litellm-mcp

The default transport is stdio. To expose the server through FastMCP's Streamable HTTP transport instead:

export MCP_TRANSPORT=http
export MCP_HOST=0.0.0.0
export MCP_PORT=8000
uv run litellm-mcp

Clients can then connect to http://localhost:8000/mcp.

Using Docker

docker build -t litellm-mcp .
docker run -i --rm \
  -e LITELLM_URL=http://host.docker.internal:4000 \
  -e LITELLM_MASTER_KEY=sk-... \
  litellm-mcp

For Streamable HTTP:

docker run --rm -p 8000:8000 \
  -e LITELLM_URL=http://host.docker.internal:4000 \
  -e LITELLM_MASTER_KEY=sk-... \
  -e MCP_TRANSPORT=http \
  litellm-mcp

Claude Desktop Configuration

{
  "mcpServers": {
    "litellm": {
      "command": "",
      "args": [
        "--directory",
        "",
        "run",
        "litellm-mcp"
      ],
      "env": {
        "LITELLM_URL": "http://localhost:4000",
        "LITELLM_MASTER_KEY": "sk-..."
      }
    }
  }
}

Tool Names

Tool names are generated from LiteLLM OpenAPI operationId values and are limited to 64 characters. A numeric suffix is added within that limit when shortening produces a duplicate name.

Examples from LiteLLM v1.84.8:

| LiteLLM Endpoint | MCP Tool | |------------------|----------| | GET /health/readiness | health_readiness_health_readiness_get | | GET /key/list | list_keys_key_list_get | | GET /key/info | info_key_fn_key_info_get | | GET /spend/logs/v2 | ui_view_spend_logs_spend_logs_v2_get | | POST /chat/completions | chat_completion_chat_completions_post |

Two helper tools are always registered:

• list_litellm_openapi_tools: Search/list dynamically registered endpoint tools.
• get_litellm_openapi_tool: Show the HTTP method, path, and parameter mapping

for one endpoint tool.

Project Structure

src/litellm_mcp/
├── configs/       # Environment configuration
├── connections/   # LiteLLM HTTP client singleton
├── tools/         # OpenAPI-driven FastMCP tool registration
└── main.py        # Entry point and tool registration

Security Notes

LITELLM_MASTER_KEY can read and mutate sensitive LiteLLM state. Because this server exposes every LiteLLM endpoint in the running proxy, including mutating admin endpoints, prefer running it locally or in a trusted network.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: wklee610
• Source: wklee610/litellm-mcp
• License: MIT

Install and usage instructions live in the source repository linked above.