MCP Server Checker
Grade your tools/list against the MCP 2026-07-28 specification and the Anthropic and OpenAI directory rules, scan every string for hidden instructions, and see what your write tools offer a caller. Works for stdio, private and OAuth servers, because you paste the list.
The check runs in this page. Nothing you paste is stored or sent anywhere, and after the page has loaded it needs no network, unless you open or make a link to a public source.
Input
Accepts an array of tools, {"tools": [...]}, the full JSON-RPC answer, or an array of answers (one per page). To get it: npx @modelcontextprotocol/inspector --cli <your server command> --method tools/list.
The example is 12 tools from GitHub's official MCP server (github/github-mcp-server, MIT licence), as published on 18 November 2026.
Results
Link to public source
A GitHub file at a commit (gh:owner/repo@commit:path, or its github.com address), a file in an npm package at an exact version (npm:package@1.2.3:path) or a jsDelivr address pinned the same way. The link holds only the source and a hash of the list. Whoever opens it fetches the file from GitHub, npm or jsDelivr in their own browser, and the page checks it against the hash.
What the fixed tool list changes
The fixed list carries only changes that a tool list can make safely: removed invisible characters, a root type: "object", titles, explicit hints and additionalProperties: false. Items marked check are inferred from tool names and specification defaults; confirm each one matches what the tool does before you ship it. Descriptions, renames, splits and new arguments need your judgement and are not made for you.
Findings by tool
Each finding names its check, its severity and the rule it comes from. Open a finding for the exact rule text, its source, and a fix. Blocking means a directory or the specification says the tool fails; attention means it is likely to be questioned or a SHOULD is missed; advice is design guidance that affects only the quality grade.
Whole list
| Tool | Findings |
|---|
Write-surface design
What the tool list offers a caller for safe writes: a retry key on creates and sends, a version check on updates and deletes, a guard on deletes. Tools are classified by name and arguments, so treat each row as a prompt to check, not a finding. It is not a test: a server with every guard can still lose a write, and a backend can protect a tool that shows none.
| Write tool | Read as | Found | Missing | Notes |
|---|
MWS 0.9 records
One record per requirement and tool, as the MCP Write-Safety Requirements ask of any tester. There is no total and no grade here (MWS 0.9, section 5.2). A tool list can decide one requirement, MWS-ANN-4.
| Requirement | Tool | Verdict | Detail |
|---|
Compare with an earlier tool list
Paste the list you reviewed before. The comparison uses a SHA-256 hash of each complete tool definition, as the OWASP MCP cheat sheet advises, and flags changes that let a client run a tool with less confirmation, such as a flipped readOnlyHint. Nothing is stored; keep the old list yourself.
| Tool | Change | Risk |
|---|
Not tested here
These need real calls against a test environment. The free mcp-writecheck GitHub Action runs the MWS tests in your own CI.
| What can go wrong | Requirements |
|---|
How the grades work
Every check, with its rule text and source: all checks. The grades cover these static rules only: they say nothing about how the tools behave, and a reviewer may decide differently.