# Put OAuth in front of your own MCP server or API so agents can sign in

> Put OAuth in front of your own MCP server or API so agents can sign in: 0 companies fully do it, 10 with a limit, 16 carry the tag. Every listing in Agent auth & delegated access with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-auth.protect-own-mcp/
- Markdown: https://www.anchorterminal.com/niches/agent-auth.protect-own-mcp/index.md (~1,600 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-auth.protect-own-mcp/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-auth.protect-own-mcp/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 10 | 16 | 22 | 10 |

## What counts as done

Claude Code, Cursor and one headless client each complete discovery, registration (DCR or CIMD), consent and a tool call against the server with no hand configuration, and a token minted for another resource is refused.

How agents ask for it: add auth to my MCP server; let Claude and Cursor users log in to our MCP; make our API usable by agents with OAuth

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Aembit (Aembit, Inc.)**: MCP Authorization Server takes DCR and CIMD client IDs, ignores the port on localhost and 127.0.0.1 redirects and checks RFC 8707 audience. Claude Code is a documented client; the audience must match the server URL precisely, and no headless client is shown.
- **Amazon Bedrock AgentCore Identity (Amazon Web Services, Inc.)**: Gateway inbound auth validates JWTs from your own IdP against allowed audience, clients and scopes and advertises protected resource metadata. AgentCore is not the authorisation server, so DCR, CIMD and loopback redirects depend on that IdP.
- **Arcade.dev (Arcade AI, Inc.)**: Arcade Deploy secures a hosted MCP server, and gateways connect Claude Code and Cursor from a URL alone; the resource-server library checks expected audiences against your own authorisation server. DCR, CIMD, loopback rules and refusing another resource's token are not documented.
- **Auth0 for AI Agents (Token Vault) (Okta, Inc.)**: Auth0 registers MCP clients by DCR or CIMD and honours RFC 8707 resource once the compatibility profile is on, but audience wins when both are sent and DCR clients need tenant-wide default permissions. Loopback redirect rules for DCR clients were not found.
- **Cloudflare Workers OAuth Provider (Cloudflare, Inc.)**: Supports CIMD and RFC 7591 DCR, loopback redirects with RFC 8252 port handling, RFC 9728 metadata and invalid_target for other resources. Trailing slash in resource stays strict, the mismatch that broke Claude Code in issue 133, and the consent page is the developer's to build.
- **Descope Agentic Identity Hub (Descope, Inc.)**: An MCP Server Resource puts its URL in aud, and the project can enable CIMD (with approved domains), DCR and pre-registration, with a hosted consent flow. Loopback redirect rules and a trailing-slash policy are not documented, and no first-hand account was found.
- **Keycard (Keycard Labs, Inc.)**: Registering the server as a Resource gives RFC 9728 metadata, audience-bound verifiers in four languages, and documented Claude Code and Cursor sign-in. The identifier must match the endpoint precisely, and how MCP clients register (DCR or CIMD) is not stated in the pages read.
- **Scalekit AgentKit (ScaleKit, Inc.)**: Scalekit's separate MCP Auth product works as the authorisation server with DCR for Claude Desktop, VS Code and Cursor and a per-server resource ID as audience. CIMD and loopback redirect rules are not mentioned in its docs.
- **Stytch Connected Apps (Twilio Inc.)**: Stytch registers MCP clients by DCR or CIMD (beta) and lets third-party public clients use loopback redirects. Tokens carry the project ID as audience, so another resource in the project is not refused, and the localhost issue that broke Claude Code (acc-24) is still open.
- **WorkOS Pipes and Agents (WorkOS, Inc.)**: AuthKit works as the MCP authorisation server with CIMD (off by default), optional DCR and validated RFC 8707 Resource Indicators. Loopback rules are not stated, and first-hand accounts report re-authentication blocked during an upstream outage and Claude Code not refreshing AuthKit tokens.

- not this job: Clerk MCP server (Clerk, Inc.): This listing is Clerk's own hosted MCP server for managing a Clerk account, not an authorisation server for the reader's MCP server.
- not this job: Microsoft Entra Agent ID (Microsoft Corporation): Agent ID makes agents confidential clients with no interactive /authorize flow; it is not an MCP authorisation server, and no DCR or CIMD is documented. A first-hand account of Entra fronting an MCP server for Claude saw a daily Connection expired.
- not this job: Nango (Nango Inc): Nango is an OAuth client to other services and MCP servers; it does not act as an authorisation server in front of your own MCP server.
- not this job: Vercel Connect (Vercel Inc.): Connect is an outbound broker that gets tokens for providers and MCP services; it is not an authorisation server in front of your own MCP server.

## What stops agents finishing

- authorisation server refuses http://localhost or ephemeral-port loopback redirects
- client and server disagree on CIMD versus DCR and the failure happens in the user's browser
- DCR client_id expires and the client shows a stuck reauth dialog
- authorisation server outage blocks every re-authentication

## Around it

Before: code.deploy-server. After: agent-observability.trace-calls.

## Where this comes from

Weight 22 of 100 in Agent auth & delegated access: 10 of 33 accounts, mostly GitHub issues on Cloudflare Workers OAuth Provider, Stytch and Claude Code from builders of MCP servers

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-auth.protect-own-mcp.json, or the MCP tool get_use_case with id agent-auth.protect-own-mcp. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
