Presso Network
Note10 of 14

Building on MCP

What does Claude's connector directory check before listing a server?

Claude's connector directory checks tool annotations, separate read and write tools, descriptions that do not instruct Claude, and its own OAuth sign-in. Here is each rule, cited, the easy mistakes, and what we found on our own three servers.

Before it lists an MCP server, Claude’s connector directory checks four things: that every tool has a title and a readOnlyHint or destructiveHint, that read and write are separate tools, that descriptions describe rather than instruct, and that Claude’s own OAuth client can sign in. Then an automated scan checks policy, and some servers get a human review.

Most of that is written down. It is spread across four pages, and the parts that trip people up are one line each.

What do the submission portal and the review check?

The Submit a connector page (opens in a new tab) (re-read on 30/10/2026) sets the requirements every connector must meet: Anthropic’s security standards, a title and the applicable readOnlyHint or destructiveHint on every tool, OAuth 2.0 for authenticated services, and clear setup and usage instructions.

The portal itself checks some of that for you. Its Tools step syncs your tools, prompts and resources from the live server, groups them as read-only or write by their annotations, and flags any tool missing a title or annotations. The page says to fix those on the server before you submit. It also asks for a documentation URL, a privacy policy URL, a support contact, an icon, test credentials for reviewers, and seven compliance acknowledgments.

The pre-submission checklist (opens in a new tab) (re-read on 30/10/2026) is what reviewers look for, organised around the most common rejection reasons:

  • Read and write in separate tools. A tool that accepts both safe methods (GET, HEAD, OPTIONS) and unsafe ones (POST, PUT, PATCH, DELETE) is rejected. Documenting the difference in the description does not count.
  • Annotations decide permissions. readOnlyHint: true lets a tool run without per-call confirmation; destructiveHint: true means Claude always prompts.
  • Tool names of 64 characters or fewer. The MCP spec only says names SHOULD be 1 to 128 characters, so the directory is stricter.
  • Narrow, accurate descriptions that match what the tool actually does.
  • No prompt-injection patterns. Descriptions are rejected if they tell Claude to call tools the user did not ask for, interfere with other tools, fetch instructions from elsewhere, hide encoded text, or promote products.
  • Functional quality. Valid input gets a successful response, bad input gets an actionable error rather than a bare “Internal Server Error”, and responses are sized for the task.
  • Scope. The server calls your own APIs, or ones you legitimately proxy. Connectors that move money or generate images, video or audio with AI models are not accepted.

Per that page, submissions are scanned automatically for policy compliance and listed as Community connectors by default. Some are escalated to Verified review, where reviewers run a functional test of each tool.

The MCP tools specification (opens in a new tab) (revision 2025-11-25, re-read on 30/10/2026) is where the field shapes come from: title is optional in the protocol, inputSchema MUST be a valid JSON Schema object and not null, and clients MUST treat annotations as untrusted unless they come from trusted servers. The directory makes the optional fields compulsory.

What is the OAuth discovery chain Claude walks?

The Authentication for connectors page (opens in a new tab) (re-read on 30/10/2026) lists the places where Claude’s OAuth client is stricter than the MCP specification. Walked in order, this is the chain:

  1. The first request has no token, and must get a 401. Its WWW-Authenticate header should carry resource_metadata pointing at your protected resource metadata. Claude ignores that header on a 200. With no pointer, Claude probes /.well-known/oauth-protected-resource/<your-mcp-path> and then /.well-known/oauth-protected-resource, which only works if your platform serves those paths.
  2. Claude reads the protected resource metadata. Its resource field must equal the server URL exactly as the user enters it, path included.
  3. Claude takes the first entry in authorization_servers and only that one. It does not fall back to later entries.
  4. Claude fetches the authorisation server’s metadata, RFC 8414 or OpenID Connect Discovery, from its /.well-known/ paths. That host must be reachable from Anthropic’s published egress range, so a firewall in front of your identity provider can break sign-in while the MCP server itself works.
  5. Claude registers itself. It uses a Client ID Metadata Document (CIMD) only when the metadata advertises both "client_id_metadata_document_supported": true and "none" in token_endpoint_auth_methods_supported. Otherwise it falls back to Dynamic Client Registration (DCR), which needs a registration_endpoint. The page recommends CIMD or Anthropic-held credentials for high-traffic servers, because DCR registers a new client on every fresh connection.
  6. Claude sends the user to sign in with PKCE S256, every time. Your authorisation server must support it, and the MCP spec requires it to advertise "code_challenge_methods_supported": ["S256"].
  7. Claude picks scopes. It uses the scope in your 401 if there is one, otherwise scopes_supported from your resource metadata, and appends offline_access only when your authorisation server’s metadata lists it. No offline_access, no refresh token.
  8. The user comes back to a redirect URI. https://claude.ai/api/mcp/auth_callback for the hosted apps, and a loopback redirect on any port for Claude Code, matched for both localhost and 127.0.0.1.
  9. Claude exchanges the code at your token endpoint, form-encoded, not JSON. Discovery, registration and token endpoints get 10 seconds; refresh gets 30. A dead refresh token must get invalid_grant, and public clients should get rotated refresh tokens.

We wrote about the resource-server half of this, the audience check that refuses tokens made for someone else, in MCP OAuth 2.1: your server is the bouncer.

Which mistakes are easy to make?

None of these is hard to fix. They are hard to notice, because the server works when you test it by hand.

  • resource that is nearly the server URL. A trailing slash, a missing /mcp, or http where the user typed https. The authentication page says the value must match exactly, including any path, and “nearly” fails.
  • Missing annotations or titles. A write tool with no destructiveHint is the common one. The portal flags it, and the review criteria tie the hints to whether Claude asks before a tool runs.
  • A catch-all request tool. An api_request tool with a method parameter is the review page’s own example of what gets rejected. Split it into a read tool and write tools, ideally create, update and delete.
  • Descriptions that instruct Claude. “Always call this tool first” reads like help to the author and like prompt injection to a reviewer. Say what the tool does and stop.

What can a script not check?

A script can read your metadata and your tool list. It cannot judge whether the server is any good. These stay with a person:

  • Test credentials for a fully populated account, entered in the portal for reviewers only.
  • Running every tool yourself, as a custom connector in Claude or through MCP Inspector. The portal’s Test & launch step asks you to confirm it.
  • Redirect URIs registered at your authorisation server, which no public endpoint reveals.
  • Whether each tool returns a useful result, an actionable error, and a response sized for the task.
  • Whether the server calls your own APIs and its domain matches your service.
  • The listing: name up to 100 characters, one-liner up to 200, description up to 2,000, one to five categories, and a URL slug that is permanent once published.
  • The directory terms and policy, and the seven compliance acknowledgments.

What did we find on our own three servers?

We ran our checker, connector-preflight, against the three MCP servers we run, on 25/10/2026, and again on 30/10/2026 with the same results.

Server What the checks found
Corey (mcp.getcorey.ai) Passes everything a script checks. One tool, report_gap, has three parameters with no description.
Presso Homes (homes.pressonetwork.com) Registration is DCR only, with no CIMD. No offline_access in scopes_supported, so Claude asks for no refresh token. Eight parameters across seven tools have no description. It is an MCP App, so it needs carousel screenshots.
Presso Events (mcp.presso.events) OAuth discovery passes end to end: a 401 with resource_metadata, resource metadata, authorisation server metadata, S256, registration and offline_access. Its tools need a signed-in token to check, so they are still to do.

Two caveats, stated plainly. Undescribed parameters are not a directory rule; Anthropic’s pages do not ask for them, and the checker reports them as a note, because Claude fills a tool’s inputs from that metadata alone. And DCR is allowed; the authentication page only recommends CIMD for servers expecting high traffic.

The carousel is the one item no script can produce: three to five PNG screenshots at least 1000px wide, cropped to the app’s response, each with its prompt supplied separately, per the submission page.

A listing is discovery, not distribution, which is the argument of A connector directory is not a marketplace. It is still worth not being turned away for something a script could have told you.

How do you check your own server?

connector-preflight is open source under MIT, read-only (it never calls a tool, reads a resource or registers a client) and has no dependencies. The latest release is v0.1.1, from 30/10/2026. It also runs as a GitHub Action (uses: PressoNetwork/connector-preflight@v0) and as a Claude Code plugin.

npx -y github:PressoNetwork/connector-preflight https://your-server/mcp

The code and the rule sources are at github.com/PressoNetwork/connector-preflight (opens in a new tab).

Where this stops working

Required field, min 1 / 4 entries
  1. 01

    The rules move. Everything here was re-read on 30/10/2026 from Claude's docs and the MCP specification revision 2025-11-25. Read the current pages before you rely on a detail; Anthropic decides what gets listed, not this note.

  2. 02

    It covers remote MCP servers submitted as connectors. Local servers now go through plugins, and the directory no longer accepts them as MCP Bundles, so most of the OAuth section does not apply to them.

  3. 03

    Passing every check a script can run does not mean you will be listed. Reviewers judge response quality, scope and the listing itself, and a server that is correct but not useful is still not useful.

  4. 04

    Our own findings come from one tool's runs on 25/10/2026 and 30/10/2026. On one server the tools sit behind sign-in, so we have not yet checked them with a token.