foxxe-mcp

Shared MCP plumbing for the FoxxeLabs server fleet. One SDK pin, one bootstrap, thin re-export of the SDK server class — so a spec revision is one commit here, not eleven latent crash-loops.

See docs/PRD-foxxe-mcp.md for why. See docs/migration.md for how to convert a server.

Status: v0.1.0 — library built, fleet unmigrated.

The shape

from foxxe_mcp import MCPServer, serve

mcp = MCPServer("tomhas")

@mcp.tool()
def status() -> str:
    """Report the current run state."""
    ...

if __name__ == "__main__":
    serve(mcp)

MCPServer is the SDK's own class, re-exported unwrapped. The tool decorator is the SDK's, reached through the instance — foxxe-mcp never wraps the registration API (PRD NG2), so every SDK feature is available the day it ships.

serve() owns what eleven servers currently hand-roll: transport, bind address, allowed host, /health, /version, logging, graceful shutdown.

Why the name is MCPServer on an SDK that calls it FastMCP

v0.1 pins mcp>=1.29.1,<2, where the class is FastMCP. The export is named MCPServer anyway — that is the mcp 2.x name, and exporting it now means the eventual pin flip changes this library and no server code at all.

That is a claim, so it is tested: scripts/verify_sdk2.py runs the same library against mcp==2.1.1 and checks that MCPServer resolves to the renamed class, that a tool still registers, and that /health, /version and /mcp all still answer. Both majors, one codebase.

$ python scripts/verify_sdk2.py
foxxe-mcp 0.1.0 against mcp 2.1.1
  PASS  SDK major detected as 2 — mcp.server.mcpserver
  PASS  MCPServer resolves to the renamed class
  ...
All checks passed — the v1.0 pin flip is a one-line change.

The bound is enforced twice

Once at install time, in pyproject.toml, and again at boot, in check_sdk(). The second one is not redundant. PRD finding B: rialu declares no mcp at all, only an unbounded fastmcp, and fastmcp 3.x pulls mcp>=2 transitively — a declaration that binds only at install time cannot catch what a resolver did to a transitive dependency. serve() calls check_sdk() before the first tool call, so a wrong resolution fails at boot, next to its cause.

mcp out of range is fatal. fastmcp out of range warns, and is fatal under FOXXE_MCP_STRICT_SDK=1.

Install

foxxe-mcp==0.3.0

Add the fastmcp extra only if the server actually uses the third-party fastmcp package: foxxe-mcp[fastmcp]==0.3.0. Never declare mcp or fastmcp directly in a server again — that is the whole point.

Published to PyPI rather than installed as a git dependency, which is a deliberate amendment to PRD §7. The git-by-tag plan died on contact with the first real deploy: this repo is private, so pip install git+https://... inside a Docker build has no credentials and fails with exit 128. The alternatives were a build-time PAT threaded through eleven Dockerfiles, or making the repo public — which would publish fleet.toml and the internal audit along with it. PyPI keeps the repo private, publishes only the package, and makes every Dockerfile a plain pip install again. NG3 rules out a private index; this is the public one.

Fleet operations

foxxe-herd status              # poll every server's /version — who is on what
foxxe-herd redeploy            # dry run
foxxe-herd redeploy --yes      # fly deploy every migrated server

fleet.toml maps server name to Fly app name (verified against flyer apps_list, not guessed — the -foxxelabs suffix is inconsistent) and records which servers are migrated and which are still unbounded.

Environment

Development

python -m venv .venv && .venv/bin/pip install -e '.[dev]'
.venv/bin/python -m pytest