Why MCP Servers Get Rejected: Common Review Failures and How to Fix Them
The rejection reasons Anthropic and OpenAI publish cluster into a few groups: tools missing or misusing annotations, catch-all tools that mix reads and writes, descriptions that steer the model or promote a product, generic error messages, missing or incomplete privacy policies, and reviewers being unable to connect or sign in. OpenAI also lists failed test cases and returning user data your privacy policy doesn't disclose. Each has a direct fix on your server or listing.
Both Anthropic and OpenAI publish what their reviewers check. Anthropic organizes its connector checklist "around the most common rejection reasons," and OpenAI's review FAQ lists its common rejection reasons by name. This post collects them with fixes. Neither company publishes rejection statistics, so the order below is not a ranking. mcplisted is not affiliated with either company.
Quick reference
| Failure | Platform | Fix |
|---|---|---|
Tool missing title or hint annotation |
Claude | Add title and readOnlyHint or destructiveHint |
| Annotations not explicit, or not matching behavior | ChatGPT | Set all three hints to true or false honestly |
| One tool accepts both read and write HTTP methods | Claude | Split into read and write tools |
| Freeform query tool with no API reference | Claude | Name or link the target API in the description |
| Description tells the model how to behave | Both | Describe the tool only |
| Generic errors like "Internal Server Error" | Both | Return specific, actionable messages |
| Missing or incomplete privacy policy | Both | Cover collection, use, sharing, retention, contact |
| Reviewer can't connect or log in | ChatGPT | Public URL, working credentials, no MFA |
| Test cases don't produce expected results | ChatGPT | Rerun all cases; tighten expected behavior |
| Returns user data not in the privacy policy | ChatGPT | Remove fields or disclose them |
| Unsupported use case | Both | Money transfers, AI media generation (Claude), digital goods sales (ChatGPT) |
1. Annotation problems
Claude requires a title and the applicable readOnlyHint or destructiveHint on every tool, and the portal flags tools that are missing them. ChatGPT requires readOnlyHint, destructiveHint, and openWorldHint as explicit booleans, and its FAQ lists "tool hint annotations do not appear to match the tool's behavior" as a common rejection.
The usual mistake is marking a tool read-only when it writes a log, queues a job, or sends a message. OpenAI says to use readOnlyHint: false for any of those. Fix the value on the server and rescan; a justification in the form does not override the server value. See tool annotations.
2. Catch-all tools
Anthropic rejects "a single tool that accepts both safe HTTP methods ... and unsafe methods." A generic api_request tool with a method parameter fails, and documenting safe versus unsafe calls in the description does not help.
Before: api_request(method, path, body)
After: get_record(id) readOnlyHint: true
create_record(fields) destructiveHint: false
delete_record(id) destructiveHint: true
OpenAI has a matching rule: expose each model-callable operation as its own tool, and don't use a generic executor to reach operations that weren't individually reviewed.
3. Descriptions that steer the model
Anthropic rejects descriptions that tell Claude to call tools the user didn't ask for, interfere with other tools, pull instructions from external sources, hide encoded instructions, or promote products. OpenAI's "fair play" rule bans model-readable text that pushes the model to prefer your plugin. Rewrite descriptions to state what the tool does and when to use it. See tool descriptions that pass review.
4. Generic or missing error handling
Anthropic says every tool must return a successful response with valid parameters, and generic errors such as "Internal Server Error" or "Bad Request" with no detail fail review. The policy requires MCP servers to "gracefully handle errors and provide helpful feedback." Return a tool result with isError: true and a message that names the problem and the fix. See JSON-RPC errors reviewers notice.
5. Privacy policy gaps
Anthropic's docs warn that missing or incomplete privacy policies "result in immediate rejection" for local connectors, and the policy requires a privacy policy link for any software that collects user data or connects to a remote service. OpenAI requires a published policy covering categories of data, purposes, recipients, retention, and user controls, and rejects plugins that return user data the policy doesn't disclose.
6. Reviewers can't get in
The first reason in OpenAI's list is "We're unable to connect to your MCP server using the MCP URL and/or test credentials." Credentials must work from outside your company network, without MFA, SMS or email codes, and must not expire. Anthropic asks for a sandbox account with sample data, not a real customer account, when a listing goes to Verified review. See privacy policy and test accounts.
Connection failures also come from infrastructure. Anthropic's troubleshooting guide lists redirects to another host, hostnames that resolve to private IPs, IPv6-only DNS, and WAFs blocking its egress range.
7. Failed test cases (ChatGPT)
OpenAI rejects plugins when test cases don't produce correct results. Rerun all five positive and three negative cases on every surface you support, check UI for loading errors, and make sure answers don't include extra information or personal identifiers. If a case is ambiguous, rewrite it.
8. Oversized or overreaching responses
Anthropic asks for responses "reasonably sized for the task" and says servers must be "frugal with their use of tokens." OpenAI says tool responses should not include session IDs, trace IDs, request IDs, timestamps, or logging metadata unless the user's request needs them. Both ban collecting conversation data beyond what the tool needs.
9. Unsupported use cases
Anthropic does not accept connectors that transfer money or crypto, or that generate images, video, or audio with AI models (design visuals like charts and diagrams are allowed). OpenAI allows commerce only for physical goods and bans selling digital products, subscriptions, or credits through a plugin.
Fix order that saves time
- Run your tools through MCP Inspector and look at every annotation and error path.
- Split any tool that mixes reads and writes.
- Rewrite descriptions to plain behavior statements.
- Audit one real response per tool for extra fields.
- Re-read your privacy policy against those fields.
- Run a free check on the live URL, then submit.
Paste your URL and see what Claude, ChatGPT and Gemini reviewers will flag.
FAQ
Does Anthropic publish rejection rates?
No. Anthropic publishes the checklist reviewers use, organized around the most common rejection reasons, but no numbers.
Can I appeal a ChatGPT rejection?
Yes. OpenAI says to reply to the rejection email with a clear rationale and any new information. For automated tool findings there is an Appeal button in the portal.
Will documenting safe and unsafe calls in one tool pass Claude review?
No. Anthropic says documenting safe versus unsafe operations in one tool's description doesn't satisfy the requirement; they must be separate tools.
Is a generic 500 error really a rejection reason?
Anthropic says generic errors such as Internal Server Error or Bad Request with no detail fail review.
Sources
- Connector pre-submission checklist (Claude docs)
- Anthropic Software Directory Policy
- Submit a connector to the directory (Claude docs)
- Troubleshoot your connector (Claude docs)
- Remote MCP server review requirements (OpenAI docs)
- Plugin guidelines (OpenAI docs)
- Plugin submission errors (OpenAI docs)
Rules change. If something here is out of date, email hello@mcplisted.com and we'll fix it. See our editorial policy.