mcp

jwt-hack mcp runs jwt-hack as a Model Context Protocol server over stdio, so an MCP client such as Claude Desktop or Claude Code can call its JWT tools during a conversation.

jwt-hack mcp

There is no port. The client launches jwt-hack mcp as a subprocess and talks to it over stdin/stdout, so running it by hand in a terminal just waits for MCP messages and looks like it hangs. That is expected. The implementation uses the rmcp crate and speaks protocol version 2024-11-05.

Tools

The server advertises five tools. Note this is a subset of the CLI: there is no jwks, scan or server tool here.

A few behaviors worth knowing before you wire this up:

• encode requires a secret unless no_signature is true, in which case it emits an alg:none token.
• verify requires a secret. If the token is alg:none and you pass a non-empty secret, it is reported invalid rather than accepted, matching the CLI and REST surfaces.
• crack in dict mode does no file I/O; it tries jwt-hack's built-in common-secret list. brute mode is deliberately capped for interactive use: length 3 max, charset truncated to 10 characters, 100 attempts per length. Presets are az, AZ, aZ, 19, aZ19, ascii.
• payload with public_key accepts a PEM literal or a file path and forges an RS256-to-HS256 confusion token. jwk_trust is not exposed here.

Wiring it into a client

The command is the same everywhere: run jwt-hack mcp as a stdio subprocess. Make sure jwt-hack is on the PATH the client uses, or give an absolute path.

Claude Code, from the project root:

claude mcp add jwt-hack -- jwt-hack mcp

Claude Desktop, in claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "jwt-hack": {
      "command": "jwt-hack",
      "args": ["mcp"]
    }
  }
}

Any other MCP client takes the same two pieces, a command and its args. After connecting, have the client list tools; it should report decode, encode, verify, crack and payload.

Example call

An MCP tools/call for decode looks like this:

{
  "method": "tools/call",
  "params": {
    "name": "decode",
    "arguments": {"token": "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9..."}
  }
}

The server returns the result as text content, for example the decoded header, claims and algorithm. Everything runs locally as a subprocess of the client, so tokens are never sent to a network service.