infocepo-infra-mcp

MCP server for infocepo infrastructure services. The HTTP server exposes the current, stateless Streamable HTTP transport at /mcp, legacy SSE at /sse, and an OpenAPI facade for diagnostics and direct tool calls.

Requirements

• Linux

• Python 3.11 or newer

• uv

The same installation works on x86_64 H100/DGX and aarch64 DGX Spark hosts. This project does not install or pin CUDA drivers.

Related MCP server: openclaw-tools-mcp

Install and run

./install.sh                         # safe to repeat; also upgrades dependencies
cp .env.example .env
# Edit .env, then export it into the process environment:
set -a; source .env; set +a
./run.sh 0.0.0.0 8085               # omit PORT to select a free port

The virtual environment is stored at ~/venv/infocepo-infra-mcp, or at $VENV_DIR when that variable is set. run.sh stays in the foreground and handles termination normally. It starts infocepo_mcp.sse_server, not the stdio entry point.

Confirm that the server is running:

curl http://localhost:8085/health

When API_TOKEN is configured, connect an MCP client with either supported authentication header:

URL: http://localhost:8085/mcp
Authorization: Bearer <API_TOKEN>

X-API-Key: <API_TOKEN> is accepted as an alternative. The health endpoint is always public. If API_TOKEN is unset, all endpoints are public; this is useful for local development but is not recommended on an exposed interface.

systemd example

Because .env is loaded by the Python process but shell variables in it are not automatically exported, use systemd's EnvironmentFile support:

[Service]
WorkingDirectory=/path/to/infocepo-infra-mcp
EnvironmentFile=/path/to/infocepo-infra-mcp/.env
ExecStart=/bin/bash /path/to/infocepo-infra-mcp/run.sh 127.0.0.1 8085
Restart=on-failure

Configuration

The two API keys have different purposes:

• API_TOKEN optionally protects this server's MCP, SSE, and OpenAPI routes.

• INFOCEPO_API_KEY authenticates calls from tools to the upstream infocepo LLM, audio, image, and diarization services. Tools that do not use those services can run without it.

Additional credentials enable service-specific tools:

The credentials JSON file accepts api_key, chroma_token, s3_access_key, s3_secret_key, registry_user, and registry_password keys. Environment variables are loaded first; values present in the JSON file override them.

HTTP endpoints

Except for /health and CORS preflight requests, all endpoints require API_TOKEN when it is configured.

Tools

The server currently publishes 19 tools:

The four discovery tools all use the same MediaWiki API. A failure reported by all four therefore normally identifies one shared wiki upstream/path problem; it does not show that the LLM, audio, image, ChromaDB, registry, or S3 tools are unavailable. Discovery responses include an upstream object containing the failing URL, HTTP status when available, and failure kind.

Open WebUI and MCPO

MCPO operations have names such as call_infra_list_services. The call_ prefix belongs to MCPO's generated OpenAPI operation and is not part of the MCP tool name (infra_list_services).

Start the provided bridge with:

docker compose -f docker-compose.mcpo.yml up -d
docker compose -f docker-compose.mcpo.yml logs -f mcpo

The sample compose file exposes MCPO at http://localhost:8000; its OpenAPI document is at /openapi.json and Swagger UI is at /docs. The checked-in mcpo-config/config.json connects MCPO to http://host.docker.internal:8085/mcp. If the MCP server has API_TOKEN set, add the same bearer token to that file's headers object, for example:

"headers": {
  "Authorization": "Bearer replace-me"
}

MCPO_API_KEY is conceptually separate from API_TOKEN: it protects MCPO's generated OpenAPI API and is configured on the MCPO process/client. The sample compose file currently passes its MCPO key directly with --api-key.

Check each layer independently:

curl http://localhost:8085/health
curl -H "Authorization: Bearer $API_TOKEN" http://localhost:8085/openapi.json
curl -H "Authorization: Bearer $MCPO_API_KEY" http://localhost:8000/openapi.json
curl "https://infocepo.com/wiki/api.php?action=query&titles=Main_Page&format=json"

Other entry points

The package also installs infocepo-mcp, the stdio transport implemented by infocepo_mcp.server, and infocepo-mcp-http, an alias for the HTTP server. For normal network operation, prefer run.sh or infocepo-mcp-http.

Tests

uv run pytest