MCP server in HTTP mode for remote deployments

How to run the Omnifox MCP server as an HTTP service for Composio, Railway, Fly, or Kubernetes, with its two real endpoints.

Aug 19, 2026

Watch this step in the video · Test the MCP server with the Inspector

Besides the stdio mode used by desktop clients, the Omnifox MCP server can run as an HTTP service. That is how you use it from hosted platforms like Composio or from your own infrastructure (Railway, Fly, Kubernetes, any container).

Starting it

MCP_SERVER_MODE=http \
PORT=8080 \
OMNIFOX_API_KEY=tu-token-aqui \
npx @omnifox/mcp-server

Endpoints

  • POST /mcp — all MCP traffic, over stateless Streamable HTTP transport: each request creates its own server and its own transport, and both are destroyed when the response closes. That lets you scale horizontally without sticky sessions.
  • GET /health — liveness probe. Returns 200 with { "status": "ok", ... }, the service name, and a timestamp.

GET /mcp and DELETE /mcp respond 405: they are not used in stateless mode.

The server accepts CORS from any origin and bodies up to 4 MB.

Registering it in Composio

In Composio's MCP integration, register your deployment's public URL ending in /mcp, for example https://mcp.tuempresa.com/mcp.

Security warning

The server does not authenticate the caller: the Omnifox token lives in its environment variable, the MCP client does not supply it. In other words, anyone who can reach your POST /mcp operates on your workspace with those permissions.

If you publish it on the internet, put it behind something that controls access: a private network, an authenticating proxy, an IP allowlist, or the access control of the platform hosting it. Do not leave it exposed unprotected.

Verify it

curl -s https://tu-host.ejemplo.com/health

A 200 response with "status": "ok" confirms the process is alive. Keep in mind that /health responds the same even if the Omnifox token is invalid: it only tests the process, not the credentials.

If something does not work

  • The process starts and dies immediately: OMNIFOX_API_KEY is missing; the server refuses to start without it.
  • 405 when connecting: your client is doing GET /mcp; stateless mode only accepts POST.
  • Timeouts with large payloads: raise OMNIFOX_TIMEOUT_MS and check the time limit of whatever proxy sits in front.
  • 413: the body exceeds 4 MB.

Tips

  • Use GET /health as a liveness probe in Kubernetes or in your load balancer.
  • One deployment per workspace, each with its own token: it is simpler to audit and to revoke.

Related articles

Was this helpful?

Related articles

Message us on WhatsApp