FastAPI backend template with MCP Streamable HTTP.

Description

This repository is a small backend template for services that need both:

• human/API-client access through REST endpoints
• agent access through MCP (Model Context Protocol) tools, resources, and prompts

The app exposes one mock domain through both surfaces:

• REST: /health, /projects, /projects/{project_id}, ticket triage, release notes
• MCP: /mcp using Streamable HTTP, with matching project/ticket tools and resources

The mock data is intentionally in memory. Add a database only after the REST and MCP boundaries are clear.

MCP Status

This template targets the MCP Python SDK v2 beta and the 2026-07-28 release candidate direction. That means:

• Streamable HTTP is the primary transport.
• /mcp is stateless HTTP in this sample.
• The SDK version is pinned exactly because v2 is still a pre-release.
• This is a template and learning sample, not a production MCP compatibility claim.

Auth Discovery

This sample does not implement OAuth or OpenID Connect discovery. Some MCP clients may probe .well-known URLs before calling /mcp, such as:

• /.well-known/oauth-protected-resource/mcp
• /mcp/.well-known/oauth-protected-resource
• /.well-known/oauth-authorization-server/mcp
• /.well-known/openid-configuration/mcp

Those requests can return 404 Not Found in this local sample. That is expected as long as JSON-RPC requests to /mcp return 200 OK or 202 Accepted.

To turn this into an authenticated HTTP MCP server, add:

• OAuth 2.0 Protected Resource Metadata for the MCP resource server.
• Authorization Server Metadata or a documented external authorization server.
• WWW-Authenticate headers on 401 Unauthorized responses that point clients to the protected resource metadata.
• Bearer token validation on every MCP HTTP request, including audience checks so tokens are scoped to this MCP server.
• Tests for anonymous access, invalid tokens, valid tokens, metadata discovery, and token audience failures.

Layout

app/
  main.py              # FastAPI app and /mcp route wiring
  mcp_server.py        # MCP tools, resources, prompts
  schemas.py           # Pydantic request/response models
  config.py            # pydantic-settings config
  internal/
    catalog.py         # mock domain logic
  routers/
    health.py
    projects.py
  tests/

Run Locally

cd app
uv sync --python 3.14 --prerelease allow
uv run uvicorn main:app --reload --host 127.0.0.1 --port 8001

Open:

• REST docs: http://127.0.0.1:8001/docs
• MCP endpoint: http://127.0.0.1:8001/mcp

Checks

Run the public checks directly with uv:

cd app
uv run ruff check .
uv run ruff format --check .
uv run mypy .
uv run pytest
uv audit

uv audit is included as the modern dependency-health command to run against the lockfile. Keep it visible, but treat upstream advisory availability and the current uv release status as operational context.

Key Notes

• Keep REST/MCP layers thin: validate transport input, call internal logic, return typed models.
• Be cautious when you make MCP for irreversible or destructive actions.
• Add auth, database, background work, and external clients as separate layers.