MCP Directory Submission Checklist: Claude vs ChatGPT vs Gemini Compared
Claude and ChatGPT both review remote MCP servers before listing them, but they ask for different things: Claude needs a title plus readOnlyHint or destructiveHint on every tool and OAuth 2.0 for authenticated services, while ChatGPT needs explicit readOnlyHint, destructiveHint, and openWorldHint booleans, domain verification, five positive and three negative test cases, and a demo video. Gemini CLI has no review queue: its gallery indexes public GitHub repos tagged gemini-cli-extension that pass validation.
If you want your MCP server listed in more than one place, it helps to see the requirements side by side. This table summarizes what each platform publishes as of November 2026. It is a summary, not a substitute for the official pages linked below, and mcplisted is not affiliated with Anthropic, OpenAI, or Google.
The big comparison table
| Requirement | Claude Connectors Directory | ChatGPT directory (plugins) | Gemini CLI extension gallery |
|---|---|---|---|
| Where you submit | claude.ai/directory/manage, MCP connector |
platform.openai.com/plugins, plugin ZIP |
No submission; public GitHub repo with gemini-cli-extension topic |
| Who can submit | Paid Claude plan; Owner on Team and Enterprise | Org owner or member with apps write permission; verified identity | Anyone with a public repo |
| Server location | Remote, https:// URL |
Remote, stable public HTTPS | Any server the extension config can launch or reach |
| Local servers | Not as a connector; include in a plugin | Desktop-only plugins, approved partners only | Yes, via command |
| Transport | Streamable HTTP; legacy HTTP+SSE still supported | Streamable HTTP | stdio, SSE (url), Streamable HTTP (httpUrl) |
| Auth options | OAuth 2.0 (DCR, CIMD, Anthropic-held client), none, static headers in beta | None, OAuth, or mixed; API key not supported in portal | OAuth discovery, headers, Google credentials |
| Required annotations | title + readOnlyHint or destructiveHint |
readOnlyHint, destructiveHint, openWorldHint as explicit booleans |
Not published as a listing rule |
| Tool name limit | 64 characters | Unique, plain-language names | Sanitized; truncated past 63 characters |
| Listing name limit | 100 characters | 30 characters | name in gemini-extension.json |
| Short description | One-liner, 200 characters | 30 characters | description in manifest |
| Long description | 2,000 characters | 4,000 characters | Not published |
| Privacy policy | Required link | Required HTTPS URL, plus terms, support, website | Not published |
| Test materials | Test account only if escalated to Verified review | 5 positive + 3 negative test cases, demo video, credentials | None |
| Domain proof | Policy: verify you own or control endpoints and domains | /.well-known/openai-apps-challenge token |
None |
| Review | Automated scan, listed as Community by default | Automated checks plus review, then you publish | Automated validation by daily crawler |
| Published review time | Not fixed | Not published | Crawl is daily; check after 24 to 48 hours |
Checklist that covers all three
Work through this once and most of each platform's requirements are covered.
Server and transport
- Serve Streamable HTTP on one stable HTTPS path such as
/mcp - No redirect from the registered URL to another host
- Hostname resolves to a public IPv4 address (Claude connectors are IPv4-only)
- No WAF rule blocking Anthropic's or OpenAI's published egress ranges
See Streamable HTTP vs SSE and how to host a remote MCP server.
Tools
- Every tool has a
title - Every tool sets
readOnlyHint,destructiveHint, andopenWorldHintexplicitly, which satisfies both Claude and ChatGPT - Names are 64 characters or fewer, verb-based, with no promotional words
- Reads and writes live in separate tools
- Descriptions say what the tool does and when to use it, with no instructions to the model about other tools
"annotations": {
"title": "Delete project",
"readOnlyHint": false,
"destructiveHint": true,
"openWorldHint": false
}
Annotation rules differ in detail between platforms; the annotations guide explains how each one reads them.
Authentication
- Unauthenticated calls to protected tools get HTTP
401withWWW-Authenticate: Bearer resource_metadata="..." -
/.well-known/oauth-protected-resourcereturns valid JSON - Authorization server metadata advertises
code_challenge_methods_supported: ["S256"] - DCR (
registration_endpoint) or CIMD (client_id_metadata_document_supported: true) is available - Redirect URIs include Claude's and ChatGPT's callbacks
Details are in OAuth for remote MCP servers and Claude callback URLs.
Listing and policy
- Privacy policy that covers collection, use, sharing, retention, and contact
- Public documentation and a support contact
- Square icon and short names that fit the 30-character ChatGPT limit
- No money transfers, no AI image, video, or audio generation (Claude), no digital goods sales (ChatGPT)
Gemini CLI extension (if you want the gallery)
-
gemini-extension.jsonat the repo root withmcpServerspointing at your server - Public GitHub repo with the
gemini-cli-extensiontopic - Tagged release so the crawler picks it up
The full walkthrough is in publishing a Gemini CLI extension.
Where the platforms differ most
Test materials. ChatGPT requires exactly five positive and three negative test cases plus a video for an MCP plugin's initial review. Claude only asks for test access if a listing goes to Verified review.
Annotations. Claude requires title and the applicable hint. ChatGPT requires all three hints as explicit booleans and scans them on every update.
Who publishes. Claude lists a passing submission as Community by default. ChatGPT approves a version and you select Publish. Gemini's gallery lists your extension automatically once it is crawled.
API keys. Claude has static header auth in beta for a limited set of organizations. ChatGPT's portal has no API-key flow. Gemini CLI users can set headers in their own config. See authentication without OAuth.
How to use this checklist
Start with the strictest items: explicit annotations on every tool, OAuth discovery that answers with a real 401, and a public HTTPS endpoint that doesn't redirect. Then run a free check to see which items your live URL already passes before you open any portal.
Paste your URL and see what Claude, ChatGPT and Gemini reviewers will flag.
FAQ
Which directory is hardest to get into?
They test different things. ChatGPT asks for the most review material (test cases, video, verified identity); Claude adds Verified review for some listings; Gemini's gallery only validates the repo and manifest.
Can one server be listed on all three?
Yes, nothing in the docs prevents it. A remote Streamable HTTP server with explicit annotations and standard OAuth discovery can be submitted to Claude and ChatGPT and referenced from a Gemini CLI extension.
Do I need openWorldHint for Claude?
Claude's checklist names title, readOnlyHint, and destructiveHint. Setting openWorldHint as well does no harm and is required by ChatGPT.
Is the MCP Registry one of these directories?
No. The official MCP Registry is a separate metadata catalog. Anthropic says publishing there doesn't surface your server in Claude.
Sources
- Submit a connector to the directory (Claude docs)
- Connector pre-submission checklist (Claude docs)
- Authentication for connectors (Claude docs)
- Anthropic Software Directory Policy
- Upload and submit your plugin (OpenAI docs)
- Plugin guidelines (OpenAI docs)
- Release extensions (Gemini CLI docs)
- MCP servers with Gemini CLI
Rules change. If something here is out of date, email hello@mcplisted.com and we'll fix it. See our editorial policy.