# What is Client ID Metadata Documents?

> CIMD lets an OAuth client identify itself to an authorization server it has never met. The client's ID is a web address, and the server reads the client's details from that address instead of requiring registration first.

Explainer by Boston Identity, 2026-10-15, updated 2026-11-06. 7 min read. https://insights.bostonidentity.com/glossary/cimd

- What it is: protocol
- Defined by: IETF OAuth WG draft, not yet an RFC
- Who it's for: Teams connecting AI agents and apps to OAuth servers

## What is Client ID Metadata Documents?

OAuth is the standard way an app gets permission to call an API for a user. The app, which OAuth calls the client, needs a unique ID to use an authorization server, the server that hands out access tokens. That server also needs the app's name, logo and the addresses it may send users back to [S1 §1]. When the two have never met, nobody can register the app by hand. Dynamic client registration (DCR, RFC 7591) lets the app register itself at run time, but the draft says this is not always practical and can leave servers with registrations to manage and inactive clients to clean up [S1 §1].

With a Client ID Metadata Document (CIMD), the client describes itself instead. Its ID is an https URL, and it publishes a JSON file about itself at that URL [S1 §3]. The authorization server fetches the file when it needs the details, so the client publishes its registration once, on its own site [S1 Abstract, §1]. A server may also pre-register a client's URL rather than rely on fetching it at sign-in. The draft expects that to be common in enterprises that want to onboard each client on purpose, and it says servers have no obligation to accept unknown clients on the fly [S1 §9.2].

The Internet Engineering Task Force (IETF) OAuth working group defines it in an Internet-Draft. It is a draft, not yet an RFC [S1]. The authors say it suits clients with a stable public web address and serves apps on a developer's own machine less well [S1 §1].

Think of it like this: A contractor arrives at a front desk with no badge, only a company web address. The desk looks the address up and checks that the page names the same company before calling the host down. The page proves someone runs that address, not that the company is any good.

Not the same as: Not the same as dynamic client registration, where the client sends its details to each server and gets back an ID the server makes up. With CIMD the ID is the client's own URL, and the server comes to read the details.

## How does Client ID Metadata Documents work?

Four parties take part: the user, the OAuth client, the web host that serves the client's URL, and an authorization server that has never seen this client before. The walk-through follows a sign-in where the user approves the app.

[Figure: The authorization server learns who the client is by fetching the document at the client's own URL, and checks it before the user is asked to approve.]

1. **The client publishes its document.** The client serves a JSON document at an https URL with a path, such as https://example.com/client. It uses the fields dynamic registration already defines, and its client_id field must match the URL exactly [S1 §3, §4]. These documents are public, so the document must not name a shared-secret authentication method. The document may carry public keys, for example a jwks_uri link to its key set, but never private ones [S1 §4.1].
2. **The client checks the server supports it.** An authorization server that publishes its own metadata (RFC 8414) uses a flag, client_id_metadata_document_supported, to say it handles these documents [S1 §8]. A client that reads the flag first avoids sending the user to a server that would reject it as an invalid client [S1 §8].
3. **The client sends the user to sign in.** The client redirects the user to the authorization server with an authorization request, and the client_id in that request is the URL itself [S1 Abstract, §10.6]. Client IDs are compared as plain strings, so https://example.com/client and https://example.com:443/client are two different clients [S1 §3].
4. **The server fetches the document.** The server should fetch the document from the URL automatically [S1 §6]. Anything but a 200 OK response must be treated as an error, and redirects must not be followed [S1 §6]. Nor may the server fetch a URL that resolves to a special-use IP address, the reserved ranges listed in RFC 6890 [S1 §10.7]. The draft recommends reading no more than 5 kilobytes [S1 §10.11]. If the fetch fails, the server should abort the request [S1 §6.1]. It may cache the document, but must not cache errors [S1 §6.2].
5. **The server checks the document and asks the user.** The server is responsible for checking that the document's client_id matches the URL it fetched [S1 §4]. The redirect addresses in the document become the registered ones, and the address in the request must match one exactly, with an exception for randomly assigned ports on loopback addresses [S1 §4.2]. The consent screen should show the client_id's hostname beside the fetched name and logo [S1 §10.6]. Flows with no redirect, such as the client credentials grant, skip the redirect check [S1 §4.2].
6. **The client signs its requests with its key.** A client that can keep a private key should use it to authenticate [S1 §5]. If the document declares the private_key_jwt method and the server supports it, the server must require a signed JSON Web Token (JWT) from the client and check it with the public key found through the document [S1 §5]. If a document names no method, the server must assume none, the method with no client authentication [S1 §5]. How the client guards its private key is out of scope [S1 §10.3].

## Why does Client ID Metadata Documents matter now?

The Model Context Protocol (MCP), which connects AI agents to tools and data, said in its November 2025 revision that clients and authorization servers should support metadata documents [BI-FIND-2025-0584](https://insights.bostonidentity.com/ref/BI-FIND-2025-0584). That revision also made dynamic client registration optional, a may kept for backwards compatibility [BI-FIND-2025-0584](https://insights.bostonidentity.com/ref/BI-FIND-2025-0584). MCP's 2026-07-28 revision deprecated it and kept it only for authorization servers that do not support metadata documents [BI-FIND-2026-0190](https://insights.bostonidentity.com/ref/BI-FIND-2026-0190). The same revision set a minimum of twelve months before a deprecated feature can be removed [BI-FIND-2026-0395](https://insights.bostonidentity.com/ref/BI-FIND-2026-0395). The draft itself never mentions AI agents or MCP [BI-FIND-2026-0322](https://insights.bostonidentity.com/ref/BI-FIND-2026-0322).

Auth0 made Auth for MCP generally available on May 6, with clients registering by metadata document [BI-FIND-2026-1183](https://insights.bostonidentity.com/ref/BI-FIND-2026-1183). Okta put CIMD in beta for AI agents on September 16, and Ping's Advanced Identity Cloud shipped it on September 29 [BI-FIND-2026-0704](https://insights.bostonidentity.com/ref/BI-FIND-2026-0704). CyberArk's September MCP feature relies on dynamic registration instead [BI-FIND-2026-0704](https://insights.bostonidentity.com/ref/BI-FIND-2026-0704). In January, Duo's Medfisch called metadata documents "the future" and said Cisco Duo was building support, with no date [BI-FIND-2026-1194](https://insights.bostonidentity.com/ref/BI-FIND-2026-1194).

An October 2025 OpenID Foundation paper found that dynamic registration "creates a large number of anonymous clients", and that metadata documents fix this while remaining "still insufficient as a workload identity" [BI-FIND-2025-0057](https://insights.bostonidentity.com/ref/BI-FIND-2025-0057). Our reading of the draft agrees: the document proves that someone controls a web address, and says nothing about the software behind it [BI-FIND-2026-0319](https://insights.bostonidentity.com/ref/BI-FIND-2026-0319).

## Where does it fit?

- Compared with Dynamic Client Registration: With dynamic client registration, the client sends its details to each server and keeps the ID that server issues. With a metadata document, the client publishes its details once at its own URL. The draft says dynamic registration can leave servers with registrations to manage and inactive clients to clean up [S1 §1].
- Compared with OAuth 2.0 Protected Resource Metadata: Protected resource metadata is published by the API, not by the client: under MCP, every server must publish one naming its authorization server [BI-FIND-2026-0190](https://insights.bostonidentity.com/ref/BI-FIND-2026-0190). The authorization server, in turn, fetches the client's metadata document to get the client's details [S1 §6].
- Compared with OAuth SPIFFE Client Authentication: SPIFFE, a workload identity standard, gives a strong identity that depends on controlling the infrastructure [BI-FIND-2025-0057](https://insights.bostonidentity.com/ref/BI-FIND-2025-0057). An OAuth draft joins the two: a client known by a metadata document must send its URL as client_id, and its SPIFFE ID must match the document's spiffe_id [BI-FIND-2026-0317](https://insights.bostonidentity.com/ref/BI-FIND-2026-0317).

## What goes wrong?

- The authorization server fetches a URL a stranger chose, which opens the door to server-side request forgery (SSRF), where an attacker makes a server send requests for them. The draft says the server must not fetch special-use addresses. Loopback fetches are allowed only for a development server on the same machine, never in production [S1 §10.7]. It should also fetch only URL schemes it knows, which helps if a client puts a javascript: URL in a field [S1 §10.7].
- A server may tie redirect addresses to the client's own origin. Without that rule, a client can try to pose as a better-known one [S1 §10.2]. A client's name and logo are its own claims, so a malicious client can copy another's [S1 §10.10]. We think an attacker can host a document naming their own callback, which leaves the consent page as the last check [BI-FIND-2026-0947](https://insights.bostonidentity.com/ref/BI-FIND-2026-0947). In one study, at least 62 of 119 consent pages did not show where the sign-in code would go [BI-FIND-2026-0947](https://insights.bostonidentity.com/ref/BI-FIND-2026-0947).
- The URL is the client's identity. A client that changes it becomes a new client, and if the domain expires or passes to someone else, a third party can take on the client's identity [S1 §10.4].
- The client can edit its document at any time, and the server may revoke tokens or ask users to consent again when keys or redirect addresses change [S1 §10.5, §10.5.1].
- An app on a developer's own machine has no public URL. An optional appendix lets the authorization server hand out client URLs for such apps, and says those clients should not get the same trust [S1 Appendix A].
- The timing of the server's fetches can tell the client's host when users are signing in [S1 §11.1].

## What should a security leader ask?

1. Which of our authorization servers accept a URL as a client_id today, and do they require redirect addresses on the client's own hostname?
2. When our server fetches a client's metadata document, what stops it from reaching internal or loopback addresses, and when was that last tested?
3. For each agent or app we publish with a client_id URL, who owns the domain, and when does it lapse?
4. Do our consent screens show the client's hostname and the address the sign-in code will be sent to?
5. Should our authorization servers accept any client URL at sign-in, or only URLs we have pre-registered?

## Sources

- [S1] [OAuth Client ID Metadata Document (draft-ietf-oauth-client-id-metadata-document)](https://datatracker.ietf.org/doc/draft-ietf-oauth-client-id-metadata-document/) (IETF OAuth Working Group (A. Parecki, E. Smith), Internet-Draft, revision -03, 2 November 2026; adopted by the working group, not yet an RFC)

## Evidence

- [BI-FIND-2025-0584](https://insights.bostonidentity.com/ref/BI-FIND-2025-0584): In the November 2025 revision, MCP changes how a client gets its identity. Authorization servers and MCP clients SHOULD support Client ID Metadata Documents, an IETF draft cited at revision -00 in which the client_id is an HTTPS URL pointing at a document about the client; the page calls this the case of client and server with "no prior relationship (most common)". Dynamic Client Registration drops to a MAY, kept "for backwards compatibility". A client that supports every option SHOULD use pre-registered credentials first, then a metadata document if the authorization server advertises support, then dynamic registration if it offers a registration endpoint, and only then ask the user for client details. Read against our copies of the neighbouring revisions, this is the middle step: in June 2025 dynamic registration was the SHOULD and metadata documents were absent, and by July 2026 the page called dynamic registration deprecated. (from "MCP's November 2025 authorization spec puts metadata documents ahead of dynamic client registration")
- [BI-FIND-2026-0190](https://insights.bostonidentity.com/ref/BI-FIND-2026-0190): We had this from WorkOS's reading of the changelog; the page confirms it first-hand. Every MCP server MUST publish protected-resource metadata naming its authorization server, and that server MUST expose its own metadata or OpenID discovery. A client obtains its identity by a Client ID Metadata Document, by pre-registration, or by dynamic client registration, and the last is now the fallback: 'Dynamic Client Registration is deprecated and retained for backwards compatibility with authorization servers that do not support Client ID Metadata Documents.' Servers should name the scopes an operation needs in the 401 challenge. ZeroID implements both registration paths; the specification has now chosen between them, and any gateway that onboards agents by dynamic registration is building on the deprecated path. (from "MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension")
- [BI-FIND-2026-0395](https://insights.bostonidentity.com/ref/BI-FIND-2026-0395): The identity changes in the 2026-07-28 revision are filed under Minor and Deprecated, and the minor ones are the new ones. Clients 'MUST key persisted credentials by the issuer identifier, MUST NOT reuse them with a different authorization server, and MUST re-register when the authorization server changes' (SEP-2352); clients must validate a present RFC 9207 iss parameter before redeeming a code (SEP-2468); a Dynamic Client Registration request must carry an application_type (SEP-837). Above them, DCR is deprecated 'as a client registration mechanism in favor of Client ID Metadata Documents' (PR #2858) and kept for authorization servers without CIMD, which we had already confirmed first-hand on the authorization page itself; what the changelog adds is the clock, since the same revision adopts a lifecycle policy with a twelve-month minimum before a deprecated feature can be removed. None of it is marked breaking. The CIMD draft the revision points at never mentions MCP; WorkOS and the OpenID whitepaper both read DCR's anonymous client as the risk this closes. (from "MCP's July changelog: a stateless protocol and clients named by web address, but no mention of gateways")
- [BI-FIND-2026-0322](https://insights.bostonidentity.com/ref/BI-FIND-2026-0322): The draft does not say what its adopters say about it. The -03 text never names the Model Context Protocol, by that name or as MCP, and never calls a client an AI agent. It does not call dynamic client registration obsolete: registration, it says, "is not always practical in some deployments", and pre-registering a client URL is "a valid deployment pattern". How the client manages its private key is out of scope, with the attestation-based and SPIFFE client-authentication drafts named as ways it might be done. It is a working-group draft intended for the standards track, now at revision -03, and it has no implementation-status section. Our briefing on the Model Context Protocol's 2026-07-28 authorization specification found MCP deprecating dynamic registration in favour of these documents, so the tie to agents is MCP's choice. (from "OAuth's URL client ID draft tightens server rules in -03; the client is still whoever controls the URL")
- [BI-FIND-2026-1183](https://insights.bostonidentity.com/ref/BI-FIND-2026-1183): Three vendors shipped controls for MCP servers in May, each at a different point in the path, and only Auth0's carried a GA label. Auth0 made Auth for MCP generally available on May 6. It protects the MCP server itself: clients register with client ID metadata documents (CIMD), and the server trades the user's token on-behalf-of for one a downstream API accepts, both labelled GA. CyberArk's Secure MCP usage, released May 7, works from the agents' end: agents reach remote MCP servers through its Identity Broker, and disabling a server cuts off their access at once. Cloudflare's MCP server portals front third-party MCP servers, and on May 28 they gained tool renaming, with no status label. Auth0's CIMD reached GA more than four months before Okta's own CIMD for AI agents went to beta on September 16. (from "Three vendors put MCP controls in three different places in May, and only Auth0's was GA")
- [BI-FIND-2026-0704](https://insights.bostonidentity.com/ref/BI-FIND-2026-0704): Okta and Ping both shipped client ID metadata documents (CIMD), and CyberArk went the other way. With CIMD, an agent or app names itself with a URL that points to its own metadata, and needs no prior registration. Okta made CIMD a beta for AI agents and GA in preview for OIDC apps on September 16. Ping's Advanced Identity Cloud shipped it on September 29. CyberArk's September MCP feature relies on dynamic client registration (DCR) instead. Okta and Ping also changed their DCR support in the same releases, so neither has dropped it. (from "Three vendors now route MCP tool calls through an identity check, and only Cloudflare's is generally available")
- [BI-FIND-2026-1194](https://insights.bostonidentity.com/ref/BI-FIND-2026-1194): Cisco Duo added two OAuth standards to a new beta integration for the Model Context Protocol (MCP) and to its OAuth 2.1/OIDC integration, also in beta, to match the MCP specification released on November 25. Dynamic client registration (RFC 7591) lets agents register themselves as OAuth clients, and Duo says admins keep full control over the access level of clients registered that way. Resource indicators (RFC 8707) let an agent name the resource it wants so Duo issues a token valid only there, which Medfisch says ensures a token for one service "cannot be replayed or accepted by another." Client ID metadata documents are not in the beta: Medfisch calls them "the future" and says Duo is building support, with no date. (from "Cisco Duo's MCP beta lets agents register themselves; Duo calls client ID metadata documents the future")
- [BI-FIND-2025-0057](https://insights.bostonidentity.com/ref/BI-FIND-2025-0057): The paper finds every existing way of identifying an agent as a client incomplete. Dynamic client registration 'creates a large number of anonymous clients'. Client ID Metadata Documents fix that and are 'still insufficient as a workload identity'. SPIFFE gives a strong workload identity that depends on controlling the infrastructure, so it does not cross into another organization. Vendor systems such as Microsoft Entra Agent ID and Okta's are named as proprietary, with limited interoperability. In the year since, the Model Context Protocol has made metadata documents its preferred method and an IETF draft has joined them to SPIFFE. (from "OAuth handles an agent in one company, working while you watch. The OpenID paper says that is where it stops.")
- [BI-FIND-2026-0319](https://insights.bostonidentity.com/ref/BI-FIND-2026-0319): Under this draft a client's identity is a web address and whatever its owner publishes there. The client_id is an https URL, which revision -03 says "MUST contain a path component other than /", and the JSON document at that URL "MUST contain a client_id property whose value MUST match" it. The authorization server learns the client's name, redirect addresses and keys by fetching that document. Its security section says a client that changes its URL is "a brand new client with no relationship to the previous one", and that losing control of the domain would let a third party "assume the client's identity". We read this as a name anyone can look up. That is more than an anonymous registration gives a server, though it says nothing about what software sits behind the name. (from "OAuth's URL client ID draft tightens server rules in -03; the client is still whoever controls the URL")
- [BI-FIND-2026-0317](https://insights.bostonidentity.com/ref/BI-FIND-2026-0317): Revision -03 settles which identifier a SPIFFE client goes by when it also has a Client ID Metadata Document. In -02 a client using a SPIFFE certificate had to send its SPIFFE ID as client_id, so it could not also be known by a metadata-document URL. Now client_id is optional for all three forms. A client that sends one must present an SVID the server can tie to it, and for a metadata-document URL the SPIFFE ID must match the document's spiffe_id. A client that sends none is identified by its SPIFFE ID, and clients known by a metadata document must send client_id. The draft notes that dropping client_id for certificates departs from RFC 8705, the mutual-TLS standard. (from "OAuth's SPIFFE draft opens metadata URLs to all three credentials and says what a self-served key cannot prove")
- [BI-FIND-2026-0947](https://insights.bostonidentity.com/ref/BI-FIND-2026-0947): One of the authors' fixes for open registration is a move to Client ID Metadata Documents (CIMD), which they say pin client identity to a "cryptographically verifiable" document, and the MCP specification's 2026-07-28 revision did deprecate dynamic registration in favour of CIMD, keeping it only for authorization servers that lack CIMD support. We think the fix is weaker than the paper makes it sound. In the CIMD draft, which we read separately, a client's identity is control of a web address, its redirect addresses come from the document it publishes, and tying them to the client's own hostname is optional for the server. An attacker can host a document naming their own callback. What is left to stop them is the consent page telling the user who is asking, and in this study at least 62 of the 119 consent pages did not show where the code would go. (from "114 of 119 MCP servers that let outsiders register an OAuth client let them choose where sign-in codes are sent")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/cimd. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
