MCP connector review

MCP Connector Review Against Anthropic's Checklist

An independent review of your existing MCP connector against Anthropic's connector pre-submission checklist, before it goes to customers or the Claude directory. Every finding comes with reproduction steps.

$390 fixed, five business days. The fee counts toward the $1,190 MCP server build if you go ahead with one.

What this review checks

Every tool gets a pass, fail or not applicable mark, grouped by the sections of Anthropic's published checklist.

Design tools that pass review

  • Separate read and write tools. I flag catch-all tools that mix safe and unsafe HTTP methods. Reads and writes belong in separate tools, with writes split by action where practical.
  • API documentation in custom query tools. A freeform query tool's description must name its target API or link to that API's documentation.
  • Tool annotations. Every tool needs a title, read-only tools need readOnlyHint: true, and tools that modify or delete data need destructiveHint: true.
  • Tool names and descriptions. Names stay within 64 characters, and each description matches what the tool does.
  • Prompt-injection patterns. I check descriptions for hidden instructions, instructions aimed at the model, requests to pull behavior from external sources, and unrelated instructions or promotion.

Server behavior and scope

  • Functional quality. Valid calls succeed, inputs are validated, errors are actionable, responses stay a reasonable size, and the server collects no conversation or user data it doesn't need.
  • API ownership. The server calls your own first-party APIs or APIs you legitimately proxy, with a domain that matches your service.
  • Supported use case. I flag unsupported cases such as financial-asset transfers and AI-generated images, video or audio. Diagrams, charts and UI mockups are allowed.

Submission materials

  • Test account. A fully populated account with working credentials.
  • Allowed link URIs. Reviewed where your server uses ui/open-link.
  • Public documentation. Required by your publish date. During review, it can still be private.

Testing and review labels

  • MCP Inspector and Claude. I run every tool through MCP Inspector, and through Claude as a custom connector where your hosting setup allows it.
  • Community and Verified. Anthropic scans submissions automatically, lists Community connectors by default and may escalate some to Verified review. I explain what each label means. This review can't guarantee either one.

What you get

  1. 1.A checklist report mapped to the sections above.
  2. 2.Tests of every tool with valid input, bad input, denied access and an upstream error.
  3. 3.A two-user permission boundary check, testing that one user can't see the other's records through any tool or parameter.
  4. 4.Each finding with its severity, impact, reproduction steps and a proposed fix.
  5. 5.One retest within 14 days after you fix the findings.

This is an independent pre-submission review: Anthropic doesn't run, approve or certify it.

How it works

  1. Step 1

    You send the server URL or package, two populated test accounts and the tool list.

  2. Step 2

    Days 1 to 4: static checks, functional tests, permission boundary tests and sign-in checks.

  3. Step 3

    Day 5: the report and a 30-minute walkthrough.

  4. Step 4

    You fix the findings. I retest once and mark each finding fixed, open or regressed.

Price and payment

$390 up front, by invoice. Covers up to five tools and one sign-in flow. Larger connectors and extra retests need a custom quote.

The $390 review fee counts toward the $1,190 MCP server build. If the review leads to a build, you pay the remaining $800. Fixes outside a full build are quoted separately, once the report shows what they are.

FAQ

Questions

Does this review mean Anthropic approves my connector?

No. This is an independent review against Anthropic's published criteria. Anthropic decides listings and labels itself.

Is this a security audit?

No. It's a functional readiness review against the published checklist, with reproduction steps. A penetration test is a separate job.

We ran a free publishing checker. Why pay?

A checker reads your metadata. This review runs your workflows with two users, bad input and wrong permissions, writes reproduction steps and retests after your fixes.

Our connector is a local stdio server only.

I can still review it. The report says which directory rules apply once you host it.

Can you fix what you find?

I quote fixes after the report. If the right fix is a rebuild, the $390 review fee counts toward the $1,190 MCP server build.

Still have a question? Ask me directly.

Source

  • Anthropic: Connector pre-submission checklist. Fetched 5 Nov 2026. The checklist above maps to its sections: Design tools that pass review, Server behavior and scope, Gather what to include with your submission, Test before you submit, and Understand how Anthropic reviews connectors.