Self-Hosting¶

By default, timesfm-mcp runs over stdio — the MCP client starts and manages the process. For multi-user or remote deployments, run it as an HTTP server.

HTTP transport¶

timesfm-mcp --http

Default: 0.0.0.0:8000. Set PORT to change the port:

PORT=9000 timesfm-mcp --http

Point any MCP client that supports HTTP transport at http://your-host:8000.

Docker¶

FROM python:3.12-slim

RUN pip install timesfm-mcp

EXPOSE 8000

CMD ["timesfm-mcp", "--http"]

Build and run:

docker build -t timesfm-mcp .
docker run -p 8000:8000 timesfm-mcp

With TimesFM:

FROM python:3.12-slim

RUN pip install "timesfm-mcp[timesfm]"

EXPOSE 8000

CMD ["timesfm-mcp", "--http"]

Note

The TimesFM image requires ~16 GB RAM on the host. Model weights (~800 MB) are downloaded at first startup. For most deployments, the baseline-only image above is the right choice.

Force baseline only¶

On a resource-constrained host, set the backend env var:

TIMESFM_MCP_BACKEND=baseline timesfm-mcp --http

This prevents the server from even attempting to import TimesFM.

Health check¶

The HTTP server doesn't currently expose a /health endpoint. Use a simple TCP check or monitor the process.

Security note¶

The server does not implement authentication. Do not expose timesfm-mcp --http directly on a public interface without a reverse proxy with auth (e.g. nginx + basic auth or a JWT middleware). For internal use behind a VPN, the default config is fine.

Reverse proxy example (nginx)¶

server {
    listen 80;
    server_name forecast.internal;

    location / {
        proxy_pass http://127.0.0.1:8000;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
    }
}