Studying tool poisoning needs the tool description and the model's reasoning side by side, and that is what this gives you: two tiny tools with no interesting behaviour of their own, and a client that prints every step. It is a lab bench, not something to install for daily use.
A small teaching setup rather than a service: an MCP client that talks to Anthropic's API, plus an `ipinfo` server it connects to, wired together so you can watch a model choose and call a tool. The point is to have both halves under your control while studying how tool descriptions influence what the model does.
- `ipinfo` — return information about the machine's own IP address, no arguments
- `ipinfo_target_ipjson` — return information about a target IP supplied in the request parameters
- A client that prints the tool list, the request sent to the model, the model's tool selection, the server's result and the final answer — so the whole decision chain is visible in the terminal
- A `mcpservers.json` file that accepts both a spawned stdio server and an SSE endpoint, so you can point the client at your own server while experimenting
An Anthropic API key, written to `mcp_client/.env` as `ANTHROPIC_API_KEY`. Node.js v22 and the MCP TypeScript SDK; the repository ships a dev container and a Docker Compose file, and `docker compose run -it mcp_client` starts the client. The `ipinfo` server is launched over `npx` straight from the GitHub repository, which is why a `package.json` sits at the repository root. Only the first server listed in `mcpservers.json` is currently used.
Build from source — clone the repository and build it, then point your client at the binary