Solutions Use case

MCP server authorization

The Model Context Protocol tells an agent how to find a tool server and how to ask for a token. It does not decide who gets one. LumoAuth is the authorization server behind your MCP servers: it publishes the discovery document, issues short-lived tokens limited to specific tools, and lets you cut an agent off the moment you need to.

Start free MCP overview
  • Teams exposing internal tools to agents over MCP
  • Companies running a public MCP server for customers
  • Security teams asked to approve an MCP rollout
app.lumoauth.dev
MCP servers registered as protected resources, with their scopes and allowed agents

"Allow all tools" is the default, and it is the problem

Most MCP servers either run with no authentication or accept any bearer token that reaches them. Once an agent is connected it can call every tool the server has, and there is no record of which agent called what.

Writing OAuth discovery, dynamic client registration and token validation into every tool server is a lot of security code to get right, repeatedly.

How LumoAuth solves it

Register the server

Name the MCP server, give it a resource identifier, list its tools as scopes, and choose which agents and users may connect. Docs

app.lumoauth.dev
Registering an MCP server and its tool scopes

Discovery works out of the box

LumoAuth publishes the protected-resource metadata and the authorization-server metadata the MCP specification requires. An agent reads one well-known address and knows where to go. Docs

One token, one tool

A token can allow tool:search and nothing else. Scopes map to tools, so a compromised agent cannot reach what it was never granted. Tokens are pinned to the server's resource identifier. Docs

app.lumoauth.dev
MCP server detail: scopes, access policy and recent connections

Clients identify themselves with a URL

An MCP client can use an HTTPS URL as its client ID. LumoAuth fetches the metadata document at that URL, validates it and shows the client's real name on the consent screen. It is on by default, admins can switch it off, and a check tool explains any rejection. Clients that prefer the API still have dynamic registration. Docs

Test with a real token

Mint a token for any agent and scope from the portal and try the call before you ship. Docs

app.lumoauth.dev
Minting a test token for an agent and scope

Verify inside your server

The docs show how to validate the token in your MCP server, which headers to return on a 401, and how to react to a revocation signal. Docs

Setting it up

Register the MCP server

AI Agents → MCP servers → Register. Add its tools as scopes.

Decide who may connect

Pick the agents or the user roles that may obtain tokens, and whether a human must approve sensitive tools.

Add token validation to the server

Verify the signature against the published keys, check the audience and the scope. Sample code is in the guide.

Connect an agent

The agent discovers the server, asks LumoAuth for a token with the scopes it needs, and calls the tool.

Standards involved

Plain names on the left, the specification on the right. The full map is in the standards reference.

MCP authorization
Model Context Protocol authorization specification
Protected resource metadata
RFC 9728
Dynamic client registration
RFC 7591, RFC 7592
Client ID metadata documents
IETF draft
Resource indicators
RFC 8707
PKCE
RFC 7636

Standards reference on the docs site

Which plan you need

MCP authorization is on every plan. The free Developer plan includes one MCP server, Pro includes five, Business has no limit.

See all plans and add-ons

Documentation

MCP overviewdocs.lumoauth.dev/mcp Authorization flowdocs.lumoauth.dev/mcp/authorization-flow Registering a serverdocs.lumoauth.dev/mcp/registration Client ID Metadata Documentsdocs.lumoauth.dev/mcp/client-id-metadata-documents CIMD sample clientdocs.lumoauth.dev/mcp/cimd-sample Server implementation guidedocs.lumoauth.dev/mcp/server-implementation Agent and MCP in five minutesdocs.lumoauth.dev/quickstarts/ai-agent-mcp-5min

Questions

Does this work with Claude, ChatGPT and other MCP clients?
Yes. Any client that follows the MCP authorization specification discovers LumoAuth through the standard metadata and completes a normal OAuth flow.
Do MCP clients have to be registered before they can connect?
No. A client can use the URL of its own metadata document as its client ID, and LumoAuth fetches and validates that document on first use. Organization admins can turn this off under platform security settings, and the usual plan limits on users and tokens still apply.
Can a human approve a tool call?
Yes. Mark a scope as needing approval and the agent's request goes to the approvals inbox, Slack, Teams, ServiceNow or Jira before a token is issued.
What about MCP servers I do not control?
You can register them as protected resources too and issue your agents tokens for them, as long as the server validates tokens from your issuer.
How fast is revocation?
A revoked agent cannot get new tokens at once, and connected servers receive a security event so they can drop the session instead of waiting for the token to expire.

Ship passwordless sign-in this week. Put guardrails on your agents today.

The Developer plan is free for 10,000 monthly users and ten agents, with no card. Everything speaks open standards, so moving in or out is a config change.

Start free Talk to us Read the docs