Free MCP check
MCP Server Checker
Grade your tools/list against the MCP specification and the Anthropic and OpenAI directory rules.
Runs on the JSON you paste, in your browser: it works for stdio, private and OAuth servers, and nothing you paste is stored or sent.
Check a tool list
How to get your tools/list JSON
The checker grades the answer your server gives to tools/list, the request every MCP client sends to learn which tools a server offers. The quickest way to save it is the MCP Inspector’s command-line mode, which starts your server (or connects to it), sends the request and prints the answer:
npx @modelcontextprotocol/inspector --cli node build/index.js --method tools/list > tools.jsonnpx @modelcontextprotocol/inspector --cli https://example.com/mcp --transport http --method tools/list > tools.jsonPaste the file’s contents above, or open the file. The checker accepts a plain array of tools, an object with a tools array, the full JSON-RPC answer, or an array of answers when the list comes in pages. Any MCP client that can print the raw answer works just as well; nothing about your server has to be public.
What the checker grades
It runs 76 checks and grades six areas separately, A to F, so a fix in one place shows up in one grade:
- MCP specification 2026-07-28. Rules a tool list can break on its own: names, input schemas that are valid JSON Schema with an object root, unique names, header mapping, icons and metadata keys.
- Anthropic directory and OpenAI directory. The tool-level rules of each directory, with a separate “ready” flag for each: no blocking item left.
- Poisoning and hidden text. Every string and key in every tool is scanned for invisible characters, bidirectional overrides, prompt markup, instructions aimed at the model and wording that steers it away from other tools.
- Schema and description quality. Argument descriptions, enums, strict schemas, pagination and naming: design advice that changes how reliably an agent uses the tools.
- Write-surface design. What the write tools offer a caller for safe writes: a retry key on creates, a version check on updates and deletes, a confirm or dry-run guard on deletes.
Every finding links to a page for its check, with the rule quoted from its source, why it matters and how to fix it. The most common ones:
- Write tools set readOnlyHint or destructiveHint
- All three hints are explicit booleans
- Reads and writes are separate tools
- Deleting tools are not marked non-destructive
- No steering of when other tools are used
- Updates send the change, not the whole object
- All 76 checks
Or browse the checks by topic:
Claude Connectors directory requirements
Anthropic’s connector pre-submission checklist lists what reviewers reject. For the tool list itself, that means:
- every tool has a
titleand the applicable hint:readOnlyHint: trueon read-only tools anddestructiveHint: trueon tools that modify or delete data; - tool names of 64 characters or fewer;
- reads and writes in separate tools, so no catch-all request tool that takes the HTTP method as an argument;
- tools that take free-form queries or request bodies name or link the API they call;
- descriptions that say what the tool does and when to use it, without instructions to the model, hidden text or wording that interferes with other tools.
The checklist also covers server behavior the tool list cannot show, such as useful error messages and response sizes. The checker marks those as not tested rather than guessing.
ChatGPT app requirements
OpenAI’s plugin guidelines ask for readOnlyHint, destructiveHint and openWorldHint on every tool as explicit booleans, true or false, even where the specification has a default.
The guidelines also rule out a generic executor that runs whatever operation an argument names, inputs for passwords, API keys, payment cards or government IDs, requests for full conversation histories, and precise user location. Each of these has its own check.
What annotations can’t do
Annotations describe a tool; they do not make it safe. The MCP specification tells clients to treat them as untrusted unless they come from a trusted server, and a client uses them only to decide when to ask the user before a call.
Whether a write lands once, refuses a stale version and reports failures truthfully depends on what the server does. That is what the write-surface grade points at and what MCP write safety covers; the MCP Write-Safety Requirements say how to test it against a running server.
How scoring works
Each tool starts at 100 and loses points per finding by severity: blocking, attention or advice. A section’s score is the mean over its tools, less a little for whole-list findings, banded into letters. Caps keep a list that breaks a specification rule or hides instructions from looking good elsewhere.
The full method is published as the rubric (version 2.1.0, sources read on 2026-11-18). Every report and private link carries the rubric version, so a grade can always be traced to the rules it came from.
Questions
Is my JSON stored or sent anywhere?
No. The checker runs in your browser on the JSON you paste or open. Nothing is uploaded, and nothing is kept after you close the page. A private link, if you make one, carries the tool list inside the part of the URL after the #, which browsers do not send to a server and the site’s analytics strip before anything is recorded.
Does a pass mean my server will be approved?
No. A pass means the static rules that a tool list can show are met. Approval is each platform’s decision, and both review more than the tool list: the server’s behavior, its authentication and its listing.
Does it work for stdio, private or OAuth servers?
Yes. The checker never connects to your server, so it works the same for a local stdio server, a server behind a VPN, an OAuth-protected server or one that is not deployed yet. You only need the answer to tools/list.
How is this different from MCP Inspector?
MCP Inspector connects to a server so you can list and call its tools. The checker grades the list Inspector gives you: specification rules, both directories’ rules, hidden text in every string, schema quality and what the write tools offer for safe retries and updates.
Which rules does it check, and how current are they?
Every check quotes its source: the MCP specification revision 2026-07-28, Anthropic’s connector pre-submission checklist and OpenAI’s plugin guidelines, as read on 2026-11-18. Each check has its own page with the exact rule text, why it matters and how to fix it.
Can tool annotations make my server safe?
No. Annotations are hints that clients use to decide when to ask for confirmation; the specification tells clients to treat them as untrusted unless the server is trusted. Whether a write is safe depends on what the server does, which only a test against a running server shows.
When the grade is not the whole story
A good grade means the tool list follows the rules. An IsoKron audit tests what the server does with real writes: concurrency, stale data, retries and partial failure.
Request an audit or a build
Tell us which server and what you need. We reply to confirm the scope, the price and the dates.