Skip to content
IsoKron

What Claude and ChatGPT require before listing your MCP server

As of 19 November 2026, Claude’s Connectors Directory asks for a title and a read-only or destructive hint on every tool, and the ChatGPT directory asks for readOnlyHint, destructiveHint and openWorldHint as explicit true-or-false values on every tool; both also review tool names, descriptions, hidden instructions and how the server behaves when it is called.

Checked against published guidance as read on 19 November 2026. IsoKron is not affiliated with Anthropic or OpenAI, and neither company has reviewed this page. Each directory decides what it lists; meeting every row here does not guarantee a listing.

Check your server

Each table below takes one body of published guidance and goes through it row by row: what it asks, what that means for the answer your server gives to tools/list, which check of the MCP Server Checker tests it, and where we read it. Rows the checker cannot test say so.

Claude: the Connectors Directory

Anthropic lists remote MCP servers as connectors. Its pre-submission checklist is, in its own words, “organized around the most common rejection reasons”, and the tool rules come first.

Claude: the Connectors Directory: requirement, what it means for a tool list, checker check, source, date read
RequirementWhat it means for a tool listChecker checkSourceDate read
A title and a safety hint on every toolEach tool has a title and one of two hints: readOnlyHint: true when it only reads, destructiveHint: true when it changes or deletes data. The policy asks for every annotation that applies, and “the portal flags tools that are missing them”.
Reads and writes in separate toolsNo single tool takes the HTTP method as an argument and so both reads and writes. Explaining the safe and unsafe uses inside one description does not meet the rule.
Free-form query tools name the API they callA tool that accepts a path, query string or request body built by the caller names that API, or links its documentation, in the description.
Tool names of 64 characters or fewerHalf the 128 characters the MCP specification allows.
Narrow, accurate descriptionsEvery description says what the tool does and when to call it. The checker reads whether one is there and long enough to do both.

Whether a description matches what the tool does is for a reviewer with a running server.

No prompt-injection patternsNo description tells the model how to behave, steers it toward or away from other tools, sends it to fetch instructions elsewhere, or carries hidden or encoded text. The checker scans every string of every tool.
No conversation data beyond what the tool needsNo argument asks for chat history, memory, conversation summaries or the user’s files.

What the server stores or logs is not visible in a tool list.

Not an unsupported use caseNo tool moves money or other financial assets, and none generates images, video or audio with an AI model. The check is a word match on names and descriptions, so read its findings yourself.

The policy’s other exclusions, such as advertising and adult content, are not checked.

Every tool works, and fails helpfullyEach tool answers a valid call successfully, refuses invalid input with a message the caller can act on, and keeps its result to a size that fits the task.

Not tested: a tool list does not show it.

Your own API, on your own domainThe server calls APIs you own or are entitled to proxy, and its domain matches your service.

Not tested: a tool list does not show it.

A remote HTTPS server with working sign-inThe portal takes an https:// URL. Tools that act on a user’s account sit behind OAuth 2.0; a server of public data may need no sign-in.

Not tested: a tool list does not show it.

Listing materials, and a server you have testedA documentation URL, a privacy policy URL, a support contact and an icon, plus your confirmation that you ran every tool yourself before submitting.

Not tested: a tool list does not show it.

Every submission is scanned automatically and listed, by default, with the Community label. Anthropic may then move a listing to a Verified review, in which a reviewer tests each tool against the checklist. Anthropic says of the Verified label: “It isn't a security audit or a guarantee”. Read on in Publish to the directory and Connector verification.

ChatGPT: the plugin directory

OpenAI lists an MCP server as part of a plugin in the directory that ChatGPT and Codex share. Its plugin guidelines set the rules for tools, and a separate page covers what a remote MCP server needs for review.

ChatGPT: the plugin directory: requirement, what it means for a tool list, checker check, source, date read
RequirementWhat it means for a tool listChecker checkSourceDate read
Three hints on every tool, each stated outrightreadOnlyHint, destructiveHint and openWorldHint are each true or false on every tool, even where the MCP specification has a default. The guidelines ask for the “appropriate explicit boolean values”.
Hints that match what the tool doesA tool that changes state, starts a job or sends something is not read-only. One that can delete, overwrite or do something that cannot be undone is destructive. A read-only tool sets destructiveHint: false. The checker compares the hints with what each tool’s name says it does.

What a tool really does when called is for a reviewer with a running server.

Plain, unique tool namesNames are unique within the server and say what the tool does, without promotional words, internal jargon or opaque identifiers.
A real description on every toolNever empty. It covers what the tool does, when to use it and its limits and side effects, and it does not push the model toward this plugin or away from another.

Whether a description matches the tool’s behavior is for a reviewer with a running server.

One tool per operationNo generic executor that runs whichever operation an argument names. Each operation the model can call is its own tool, with its own schema and hints.
Safe to retry, or says that it is notA write tool either declares idempotentHint: true or says in its description when a repeated call repeats the effect.

Whether a repeated call really has no second effect is for a test against a running server.

Only the inputs the task needsNo argument for the whole conversation history, and none for the user’s precise location.
No restricted data as inputNo argument for a password, API key, one-time code, payment card number or government identifier. The guidelines also restrict protected health information, which the checker does not look for.
No hidden or manipulative instructionsTool metadata does not ask for secrets, try to override the platform’s instructions, conceal what a tool does, or hide instructions in encoded text.
Results limited to what was askedA tool returns only data that bears on the request: no personal data your privacy policy leaves out, no internal identifiers, no secrets.

Not tested: a tool list does not show it.

A reachable server and a working test accountA public endpoint, not a local or test one. Where tools need sign-in, reviewers get a demo account with sample data that works without a second factor.

Not tested: a tool list does not show it.

A verified publisher, a privacy policy and a support contactIdentity verification for the name you publish under, a published privacy policy, and contact details where users can reach you.

Not tested: a tool list does not show it.

Test cases for the reviewersFor a plugin with an MCP connection: five positive test cases, three negative ones and a video walkthrough.

Not tested: a tool list does not show it.

A submission may be scanned automatically, reviewed by a person, or both. Among the common rejection reasons OpenAI lists: “Tool hint annotations do not appear to match the tool’s behavior”. After a plugin is published, OpenAI fetches the tool list again from time to time, and a new or changed tool goes live only after it passes automated checks. Read on in Remote MCP server review requirements.

The MCP specification: what annotations are

Both directories build on fields that the protocol itself defines, and the protocol is careful about what those fields mean. Revision 2026-07-28 is the current one.

The MCP specification: what annotations are: requirement, what it means for a tool list, checker check, source, date read
RequirementWhat it means for a tool listChecker checkSourceDate read
Annotations are optional hints with defaultsA tool may carry title, readOnlyHint, destructiveHint, idempotentHint and openWorldHint. A hint left out takes the cautious default: not read-only, destructive, not idempotent, open world. A key spelled another way, such as destructive_hint, is not one of these fields, so the default applies.
Clients do not trust annotations from servers they do not trustA hint describes a tool. It grants nothing and proves nothing about what the tool does.

Not tested: this is a rule for clients, not for a tool list.

Tool namesOne to 128 characters, from A-Z, a-z, 0-9, underscore, hyphen and dot, unique within the server. The specification puts each of these at SHOULD.

The specification says clients must treat annotations as “untrusted unless they come from trusted servers”, and its schema calls them hints that are “not guaranteed to provide a faithful description of tool behavior”. That is why each directory reviews them. Read on in MCP specification 2026-07-28: Tools and MCP specification 2026-07-28: Schema reference.

What the checker does not test

The checker grades the tool list you paste, in your browser, and never connects to your server. That leaves out everything below, and a good grade says nothing about it.

  • What a tool does when it is called. The checker reads names, descriptions, schemas and hints. It cannot tell whether a tool marked read-only writes, whether a retry is safe, or whether a description is accurate.
  • How the server responds. Success on valid input, useful errors on invalid input, the size of results and how long a call takes all need a running server.
  • Sign-in and transport. OAuth, the HTTPS endpoint and whether a reviewer’s test account works.
  • Everything outside the tool list. API ownership, the privacy policy, documentation, support contacts, screenshots, test cases, publisher verification and each directory’s terms.
  • Policy questions. Which use cases a directory accepts, and its rules on commerce, advertising and content. The checker word-matches two unsupported use cases and nothing more.
  • What a tool returns. Whether results carry more personal data than the request needs.
  • The review itself. Each directory runs its own scans and reviews, on its own schedule, and can change its rules after the date above.

Hints in particular are claims, not proof. Whether a tool marked read-only leaves data alone, and whether a repeated call is harmless, are things only a test against a running server shows. The MCP Write-Safety Requirements (an open spec, draft 0.9, in public comment) define those tests.

Sources and dates

Every row above comes from one of these documents, each published by the organization whose rules it states. We read them on the dates shown. If a rule has changed since, the source wins.

IsoKron is not affiliated with Anthropic or OpenAI, and neither company has reviewed this page. Each directory decides what it lists; meeting every row here does not guarantee a listing. Product and company names are used only to say whose guidance a row describes.

See also the directory checks and the annotation checks, one page each, with the fix for every finding.

Check your server