MCP help server -- Docker build

Production-deployment recipe for hosting the MCP help server (the read-only introspection suite) on www.swi-prolog.org (or any other reachable host).

The container builds SWI-Prolog from swipl-devel and the MCP pack from JanWielemaker/MCP, then runs examples/swipl_help_public.pl under library(http/http_unix_daemon). That loader registers the four read-only server modules only:

• library(mcp/servers/prolog_help) -- manual, xref, pldoc
• library(mcp/servers/swipl_kernel) -- C source grep
• library(mcp/servers/source) -- file listing / reading / grep
• library(mcp/servers/git) -- log / blame / show / diff / grep

It deliberately omits target_toplevel (spawns child processes) and dev_reload (calls make/0 in-process) -- those are appropriate only in a trusted developer loop.

Build & run

make image    # build the mcp-help image
make run      # detached container on port 3410 (host)
make logs     # follow the swipl stdout
make restart  # stop + rm + run, picking up any image rebuild
make stop     # stop + rm

The container listens on TCP 3410 inside; the published port defaults to 3410 on the host and is overridable with `make run PORT=...`. Endpoints (per the MCP 2024-11-05 transport):

• POST /mcp/messages -- JSON-RPC request, JSON reply + Mcp-Session-Id header
• GET /mcp/sse -- long-lived SSE stream for notifications

Rebuilding individual layers

Bump the corresponding ENV line in Dockerfile so Docker rebuilds just that layer (and everything after it):

make update-swipl   # rebuild swipl + MCP layers
make update-mcp     # rebuild MCP layer only

Both targets sed today's date into the matching _VERSION line and then `docker build`.

Health check

HEALTHCHECK POSTs an initialize request through to `/mcp/messages`; the container is unhealthy when the dispatcher can't respond to a valid JSON-RPC request within 5 s. Health-check output goes to `/tmp/health.log` inside the container.

Behind a reverse proxy

For deployment on www.swi-prolog.org, terminate TLS in nginx and proxy `/mcp/messages and /mcp/sse to http://localhost:3410`. The SSE handler holds the connection open indefinitely with a 15 s heartbeat; configure the proxy with a matching or larger read timeout (`proxy_read_timeout 1d;` is typical) and disable response buffering for the SSE location:

location /mcp/sse {
    proxy_pass             http://localhost:3410;
    proxy_buffering        off;
    proxy_read_timeout     1d;
    proxy_http_version     1.1;
    proxy_set_header       Connection "";
}