fastmcp-kit 0.1.1

Framework for building authenticated, scoped, observable MCP servers on top of FastMCP. Tool registry, scoped bearer tokens, Prometheus + OTLP metrics, Grafana dashboards from code, multi-tenancy. `pip install fastmcp-kit` then `from mcp_toolkit import MCPToolkit`.

fastmcp-kit

0.1.0 live on PyPI. pip install fastmcp-kit then from mcp_toolkit import MCPToolkit. The distribution name (fastmcp-kit) and the Python module name (mcp_toolkit) differ — same pattern as pip install pyyaml → import yaml. PyPI's similarity rules rejected the shorter name.

A Python framework for building authenticated, scoped, observable MCP (Model Context Protocol) servers on top of FastMCP.

You bring the tools. fastmcp-kit brings:

• Tool registry — register tools with @toolkit.tool(...), group them, version them.
• Scoped auth — bearer tokens carry scopes; list_tools is filtered per caller. Tokens never see tools they can't call.
• One-click observability — Prometheus metrics auto-registered per tool, Grafana dashboards generated from the registry, docker compose up brings the whole stack live.
• Multi-tenancy — pluggable tenant resolvers (header / subdomain / token-claim). Per-tenant metric labels enforced at registration.
• Battle-tested harness — Clara architecture, mypy-strict, 80% coverage gate, bandit + pip-audit + gitleaks + CodeQL, branch protection, PR-from-fork model.

Quick taste

from mcp_toolkit import MCPToolkit

toolkit = MCPToolkit(name="my-server")

@toolkit.tool(group="weather", scopes=["read:weather"])
async def get_weather(city: str) -> dict:
    return {"city": city, "temp_c": 21.0}

@toolkit.tool(group="admin", scopes=["admin"])
async def reset_cache() -> None:
    ...

app = toolkit.build_app()  # FastAPI / FastMCP app; uvicorn-ready

Tokens minted with scopes=["read:weather"] discover and call get_weather. They cannot see reset_cache exists.

Why not just FastMCP?

FastMCP gives you the MCP wire protocol and transports. fastmcp-kit gives you the production layer above it: auth-scoped discovery, metric-per-tool, dashboards-from-code, multitenancy. You can use both — fastmcp-kit wraps FastMCP under the hood.

Install

# Just the framework
uv add fastmcp-kit

# With observability batteries (Prometheus exposition + Grafana JSON gen)
uv add 'fastmcp-kit[observability]'

# Everything
uv add 'fastmcp-kit[observability,redis,otel]'

Import name stays mcp_toolkit:

from mcp_toolkit import MCPToolkit

One-click stack

make stack-up      # Prometheus + Grafana + your server, dashboards provisioned
open http://localhost:3000  # Grafana, dashboards already populated

See deploy/observability-stack/README.md.

Deploy

Three deployment surfaces ship in this repo:

Target	Manifest	When
Knative	deploy/knative-serving.yaml	You already run Knative Serving.
Vanilla k8s (Helm)	deploy/helm/mcp-toolkit/	You run k8s without Knative.
Local dev	deploy/observability-stack/	docker compose, Prometheus + Grafana included.

Documentation

Topic	Where
AI navigation primer	CLAUDE.md
10,000 ft architecture	ARCHITECTURE.md
Tool registration	src/mcp_toolkit/domains/registry/DOMAIN.md
Auth + scopes	src/mcp_toolkit/domains/auth/DOMAIN.md
Metrics + dashboards	src/mcp_toolkit/domains/observability/DOMAIN.md
Multi-tenancy	src/mcp_toolkit/domains/tenancy/DOMAIN.md
Roadmap	docs/ROADMAP.md
Contributing	CONTRIBUTING.md
Security policy	SECURITY.md

License

MIT — see LICENSE.