mcp_auth_server
A proof-of-concept MCP server implementing OAuth 2.1 authorization with CIMD client registration and PKCE, demonstrating protected resource access and step-up authentication.
README
mcp_auth_server
A local, runnable proof of concept of the MCP Authorization spec (2025-11-25).
It demonstrates, end to end:
- an MCP server acting as a proper OAuth 2.1 resource server — publishes RFC 9728 Protected Resource Metadata, validates audience-bound access tokens, and enforces scope on a protected tool;
- a one-command MCP client that discovers the authorization server, registers via Client ID Metadata Documents (CIMD), runs the full OAuth 2.1 + PKCE handshake (including a step-up re-authorization), and calls the protected tool;
- a minimal custom OAuth 2.1 authorization server behind the MCP server.
Why a custom authorization server?
CIMD is draft-ietf-oauth-client-id-metadata-document-00. No off-the-shelf OSS or SaaS
authorization server supports it yet, so using one would still require writing a CIMD shim.
A small custom AS gives full, correct CIMD support and lets the whole thing run locally as a
single command.
Quick start
uv venv && uv pip install -e ".[dev]" # one-time setup
./scripts/run.sh # or: uv run python -m mcp_auth.demo
run.sh starts the authorization server and the resource server, then runs the client, which
prints a step-by-step trace of the handshake and the protected tool's result, e.g.:
[1] Discovery
401 challenge -> resource_metadata=...
PRM: resource=http://localhost:8001/mcp as=['http://localhost:9000/'] scopes=['mcp:connect']
AS metadata discovered at http://localhost:9000/.well-known/oauth-authorization-server
[2] Authorization (CIMD + PKCE) for base scope
[3] Call unprotected tool 'whoami'
[4] Call protected tool 'get_weather' (expect insufficient scope)
[5] Step-up re-authorization for the extra scope
[6] Retry 'get_weather' with stepped-up token
=== Success: full CIMD + OAuth 2.1 + step-up flow completed ===
Architecture
src/mcp_auth/
config.py shared ports, URLs, scopes, token TTL
keys.py authorization-server RSA signing key + JWKS export
auth_server/ custom OAuth 2.1 authorization server (Starlette)
app.py AS metadata, /authorize, /token, /jwks
cimd.py fetch + validate Client ID Metadata Documents
store.py single-use authorization-code store
tokens.py RS256 JWT minting (aud=resource) + PKCE S256 verify
resource_server/ FastMCP server as an OAuth 2.1 resource server
server.py FastMCP app + whoami + protected get_weather tool
verifier.py JWT verifier: signature, issuer, expiry, audience
client/ one-command CIMD-driven MCP client
metadata.py the client's CIMD document + a server that hosts it
flow.py 401 -> PRM -> AS discovery -> PKCE authorize -> token
run.py full journey incl. step-up
demo.py orchestrator (single command)
Default ports: authorization server :9000, resource server :8001, client CIMD host
:3000. The canonical resource URI (the token audience) is http://localhost:8001/mcp.
How the flow maps to the spec
| Spec requirement | Where |
|---|---|
| RFC 9728 Protected Resource Metadata | FastMCP, auto-served at /.well-known/oauth-protected-resource/mcp |
401 + WWW-Authenticate with resource_metadata |
FastMCP RequireAuthMiddleware |
RFC 8414 AS metadata incl. code_challenge_methods_supported, client_id_metadata_document_supported |
auth_server/app.py |
CIMD: fetch client doc, validate client_id == URL, validate redirect_uri |
auth_server/cimd.py |
Authorization code + PKCE S256 |
auth_server/app.py, client/flow.py |
RFC 8707 resource parameter in auth + token requests |
client/flow.py |
Audience-bound tokens; resource server validates aud |
auth_server/tokens.py, resource_server/verifier.py |
Scope enforcement + insufficient_scope + step-up |
resource_server/server.py, client/run.py |
Tests
uv run pytest
Covers AS/PRM metadata shape, CIMD validation (mismatched client_id, unlisted redirect URI,
missing fields, SSRF host), PKCE and token claims, audience/expiry rejection in the verifier,
and a full single-command end-to-end run.
Deliberate deviations from the spec (because this is a local PoC)
These are the conscious trade-offs that let the PoC run fully locally with one command. Each would be tightened for a real deployment:
- HTTP on
localhostinstead of HTTPS. The spec mandates HTTPS for AS endpoints and the CIMDclient_idURL; localhost redirect URIs are already spec-permitted. The MCP SDK'sAnyHttpUrlacceptshttp, so no patching is needed. - Headless auto-approved consent. The AS authenticates a fixed demo subject and issues a code without a browser prompt, so the client reads the redirect directly (no callback listener). A real AS would authenticate the user and show a consent screen.
- In-memory, process-local state (signing key, authorization codes). No persistence, no refresh-token rotation.
- SSRF guard is a localhost allowlist for CIMD fetches, sufficient for local use; a real AS needs a stricter trust policy per the CIMD draft's security considerations.
Out of scope
infra/terraform/(anticipated by.gitignore) — a local PoC needs no cloud infra.- Multi-user login UI, Dynamic Client Registration (CIMD is used instead), persistent storage.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.