cerbos-fastmcp

FastMCP middleware powered by Cerbos. Authorize every MCP tool call, prompt request, and resource query against your Cerbos policies without rewriting your FastMCP server.

Why cerbos-fastmcp

• Use the Cerbos Policy Decision Point (PDP) you already trust.
• Apply fine-grained rules to tools, prompts, and resources.
• Bring your own principal builder (sync or async).
• Configure through environment variables for easy deployment.
• Ship with an example server and matching policies.

Getting started

pip install cerbos-fastmcp

Prefer uv?

uv pip install cerbos-fastmcp

Heads up Install the Cerbos CLI locally so you can run the PDP alongside your FastMCP server during development.

Quick start

from cerbos.sdk.model import Principal
from fastmcp import FastMCP
from fastmcp.server.dependencies import AccessToken

from cerbos_fastmcp import CerbosAuthorizationMiddleware


def build_principal(token: AccessToken) -> Principal | None:
    if token is None:
        return None

    return Principal(
        id=token.claims["sub"],
        roles=token.claims.get("roles", []),
        attr={
            "department": token.claims.get("department", ""),
            "region": token.claims.get("region", ""),
        },
    )


app = FastMCP("My Cerbos-protected MCP", auth=my_auth)
app.add_middleware(
    CerbosAuthorizationMiddleware(
        principal_builder=build_principal,
        resource_kind="mcp_server",
    )
)

The middleware creates a Cerbos gRPC client using CERBOS_HOST during the FastMCP on_initialize hook, verifying connectivity before any requests are handled. Provide an AsyncCerbosClient instance if you want to manage connections yourself.

Policy model

The middleware expects a Cerbos resource policy where the kind defaults to mcp_server. Each FastMCP operation maps to an action string:

• tools/list gate the tool catalogue.
• tools/list::<name> decide if a tool is visible.
• tools/call::<name> authorize execution.
• prompts/list and resources/list cover the remaining MCP commands.

A complete sample lives in policies/mcp_tool.yaml and is reproduced in docs/policies.md. It also demonstrates schema usage and regional constraints for data access.

Configuration

Environment variables let you tweak behaviour without code changes:

Variable	Purpose
CERBOS_HOST	Cerbos PDP gRPC endpoint (host:port).
CERBOS_RESOURCE_KIND	Default resource kind (defaults to mcp_server).
CERBOS_TLS_VERIFY	true/false or a CA bundle path for TLS validation.

Example server

Run the bundled demo server and PDP in one command:

cerbos run -- uv run python -m cerbos_fastmcp.examples.server

The server listens on port 8000 and uses the policies in policies/. Import cerbos_fastmcp.examples.create_example_server() in your own tests if you need a pre-wired FastMCP instance.

Testing

Install the dev dependencies and execute the test suite inside a Cerbos context:

uv pip install '.[dev]'
cerbos run -- uv run pytest

cerbos run launches a temporary PDP, sets CERBOS_GRPC/CERBOS_HTTP, and then hands control back to pytest.

Production guidance

Local development → Install Cerbos on your workstation (see the Getting Started note) so you can run the PDP alongside FastMCP.

Production → Operate Cerbos as a managed service instance—ideally as a sidecar next to your MCP server. The Cerbos documentation covers deployment patterns, configuration, and operational best practices.

Bolt on Cerbos Hub for production control plane needs: policy distribution, CI integration, audit logs, and a collaborative policy IDE for the teams managing access to your MCP server.

Documentation

Extended guides live under docs/: installation, configuration, policy design, testing strategy, and details about the example server.

License

Apache 2.0 © Cerbos