keycardai-fastmcp

FastMCP integration for Keycard OAuth: protect FastMCP servers with Keycard authentication and run delegated OAuth 2.0 token exchange (RFC 8693) for downstream APIs.

This is the canonical home for the integration. The old name, keycardai-mcp-fastmcp, is retired; its final release (0.21.0) remains on PyPI so existing pins keep resolving.

Installation

pip install keycardai-fastmcp

Quick Start

from fastmcp import FastMCP
from keycardai.fastmcp import AccessContext, AuthProvider

auth_provider = AuthProvider(
    zone_id="abc1234",
    mcp_server_name="My Server",
    mcp_base_url="http://localhost:8000",
)

mcp = FastMCP("My Server", auth=auth_provider.get_remote_auth_provider())

@mcp.tool()
async def call_external_api(
    query: str,
    access: AccessContext = auth_provider.grant("https://api.example.com"),
):
    token = access.access("https://api.example.com").access_token
    return f"Results for {query} (token starts with {token[:8]})"

Declaring the grant as a typed parameter default injects the populated AccessContext per request; the parameter never appears in the tool's input schema. Exchange failures are recorded on the AccessContext (check access.has_errors() / access.get_errors()), never raised. Granting multiple resources is all-or-nothing: if any exchange fails, the context carries that resource's error and no tokens.

If you lint with flake8-bugbear or Ruff's B008 rule (function call in argument default), exempt your tool modules: the call-in-default is the intended spelling here, the same pattern as FastAPI's Depends. In pyproject.toml:

[tool.ruff.lint.per-file-ignores]
"src/my_server/tools/*.py" = ["B008"]

Migrating from the decorator form

The decorator form (@auth_provider.grant(...) above the tool) still works from the same object. Reading the result via ctx.get_state("keycardai") is deprecated and emits a DeprecationWarning; helpers that only hold the FastMCP Context can use await AccessContext.from_context(ctx) instead.

The warning fires once per tool, at decoration time (module import). If your test or CI setup escalates warnings to errors (-W error, filterwarnings = ["error"] in pytest config), importing a server module that still uses the old form will raise instead of warn. Either migrate the tools to the injected-parameter form, or allow this warning explicitly:

filterwarnings = ["error", 'default:Tool .* uses the grant decorator:DeprecationWarning']

Testing

Fake delegated access without patching internals:

from keycardai.fastmcp.testing import mock_access_context

with mock_access_context(access_token="fake_token"):
    ...  # grants resolve to an AccessContext serving fake_token

The bare access_token form serves the token for any resource, so it will not catch a mistyped resource URL in an access(...) call. Pass resource_tokens={...} when the test should enforce which resources the tool reads; resources outside the dict raise ResourceAccessError, matching production.

Migration from keycardai-mcp-fastmcp

The keycardai-mcp-fastmcp package is retired. Its final release, 0.21.0, stays on PyPI so existing pins keep resolving: from keycardai.mcp.integrations.fastmcp import AuthProvider emits a DeprecationWarning pointing here and returns the same class. New code should depend on keycardai-fastmcp and import from keycardai.fastmcp.