glama-mcp

glama-gateway-mcp

A Models Context Protocol (MCP) server that exposes the Glama AI gateway (https://gateway.glama.ai/v1, an OpenAI-compatible endpoint for 100+ models) as MCP tools. Standard library only, Python 3.9+, no model downloads.

Statement of need

MCP clients (Claude Desktop, Cursor, and a growing list of agent runtimes) each expect an MCP server per capability. When a research project needs to call language models across many providers — openai/..., anthropic/..., google/..., and dozens more — wiring every client to every provider is a tangle of credentials and bespoke plugins. Glama centralizes provider access behind one OpenAI-compatible API, but it does not, by itself, hand a model to an MCP client.

glama-gateway-mcp closes exactly that seam. It is a thin, dependency-free MCP server that fronts the gateway: an agent (or IDE, or personal assistant) talks MCP to this one server and gains openai/gpt-4o, anthropic/claude-2, and any other gateway model through four tools — listing models, chat completion, streaming completion, and request-status lookup. One API key, one credential, one endpoint for every connected application.

Distinguishing design choices:

• stdio MCP, implemented on the stdlib — the protocol layer uses only json, urllib and sys, so it runs anywhere Python runs and is trivial to audit.
• OpenAI-compatible request/response shapes — the gateway returns raw gateway bodies, so no field mapping is ever wrong.
• Server-side stream reassembly — MCP clients that cannot hold a live SSE stream still receive the full completion text plus usage metadata.
• Clean JSON-RPC errors — a missing key, unreachable gateway, or unknown tool yields a readable MCP error instead of a hang or silent failure.

Install

pip install .
# development:
pip install -e ".[dev]"

Requires Python 3.9+. Runtime dependencies: none.

Usage

Export your Glama key, then run the server:

export GLAMA_API_KEY="your_key_here"
export GLAMA_DEFAULT_MODEL="openai/gpt-4o"    # optional
glama-mcp

or run it as a module:

python -m glama_mcp

Connect any MCP client to the glama-mcp stdio command. The exposed tools:

Tool	Purpose
glama_list_models	list models available through the gateway
glama_chat_completion	one-shot chat completion
glama_stream_completion	streamed completion (reassembled)
glama_request_status	status of a completion request by id

Per-application wiring

Ready-to-paste mcpServers blocks for individual applications live in configs/: Celebrum, Samvit, Collabuild, S-AI, hermes-agent, teddy-techlearn, health-quest, ai-content-studio, and the portfolio site. Each block registers this server for that application's MCP client.

{
  "mcpServers": {
    "glama-gateway": {
      "command": "glama-mcp",
      "env": { "GLAMA_API_KEY": "your_key_here" }
    }
  }
}

Verifying the server by hand

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | python -m glama_mcp

Tests

python -m pytest

The suite covers the gateway client (payload shapes, SSE reassembly), protocol handshake, tool discovery and dispatch, error handling, missing-key behaviour, and a full stdio round trip.

Contributing

See CONTRIBUTING.md. Keep it dependency-free and Python 3.9 compatible; every change needs a test.

JOSS paper

Submission materials in paper/: paper.md, paper.bib, and JOSS_SUBMISSION_READINESS.md (records which gates are met and which still require calendar time).

License

MIT — see LICENSE.

AI usage disclosure

Code and documentation were drafted with generative-AI assistance and reviewed by the human maintainer, who made the design decisions. This disclosure is kept in line with the JOSS AI usage policy.