Adamant Solutions

MCP tool check

An MCP server describes its tools to the AI agent in plain text, and the agent trusts that text. A malicious or compromised server can hide instructions in it ("tool poisoning"), or change a tool after you approved it. Paste a server's tools/list to look for the known warning signs, see what the tools cost in tokens, and save hashes so you notice when something changes.

A first look, not a security audit. It finds known patterns and cannot prove a server is safe. A clean result is not a guarantee, and some findings will be false alarms. We only count anonymous visits (no cookies, no IP).

How to use this tool

  1. Get the tool list of the MCP server. Most MCP clients and inspectors can show the tools/list response; you can also send the tools/list request to the server yourself.
  2. Paste the JSON above (the whole response, the result object, or just the array of tools) and press Check tools.
  3. Read the findings, starting with the red ones. Each names the tool and the field.
  4. Press "Download hashes" and keep the file. Next time, paste it in the second box: any tool whose name, description or schema changed is reported.

Example

The classic example is a tool called add whose description says it adds two numbers but also contains a hidden block telling the agent to read ~/.ssh/id_rsa and not to tell the user. Press "Try an example" to see it reported, together with an invisible character, a duplicate tool name and an encoded blob.

What it checks, and what it does not

  • Invisible and text-direction characters (including the Unicode "tag" range) in names, descriptions and schema fields.
  • Phrases that instruct the agent: ignore earlier instructions, hide something from the user, act secretly, add a step before using the tool, call another tool.
  • Mentions of sensitive files and credentials, URLs, HTML comments and long encoded strings.
  • Duplicate tool names, very long descriptions and very long tool lists.
  • A token estimate (about four characters per token; not an official tokenizer) and a SHA-256 hash per tool.

It does not test the server, its code or what the tools actually do, and an attacker can phrase instructions in ways this list does not know. Use it as one step: review the tools yourself, run only servers you trust, and give agents the least access they need.

Privacy and security of what you paste

The tool list is processed in your browser. It is not sent to us or stored. Tool descriptions are usually not secret, but schemas may contain internal names or examples: remove anything confidential first. Ads, if shown, appear only at the bottom of the page and are loaded only after your consent choice; what you paste is never sent to us or to ad providers. See the privacy notice.

Related: MCP security review · agent-readiness check · OpenAPI security check.

Want an MCP server reviewed, or your API exposed to agents safely?

We review MCP servers and the APIs behind them, and build secure ones. See the MCP security review and the MCP packages.

Message us on LinkedIn