An agent's whole toolkit is a folder of GET URLs

Most agent stacks ship a tools framework: JSON schemas, function descriptions, a runtime that parses the LLM's tool calls and dispatches to Python. Every new tool is code.

You skip the framework. Each tool: query the database, restart the worker, send a notification, search the docs: is a hoody-curl wrapped GET URL. You list them in the system prompt as plain URLs with a one-line description each. The agent picks one, fetches it, gets a JSON response, decides what's next.

Added a tool? One URL, one description line. Removed a tool? Delete the line.

The agent never imported a library, never learned a tool-calling protocol. It just fetches URLs, which every LLM has been doing since GPT-3.

The tool registry is a list of URLs. That's the entire integration.