# Working with AI agents

noodlelab is built to be used by coding agents as much as by people. An agent
gets the same discipline a person does (units on every number, uncertainty on
every measured input, requirements checked with margins, a record of every run)
and you get work you can review instead of work you have to trust.

There are three ways in:

- **The MCP server**, `noodlelab mcp`. Add it to Claude Code, Codex CLI or any
  MCP client and the agent can find nodes, build and edit graphs, run them and
  verify the result. [The MCP tools](mcp.md) lists every tool.
- **The Python library**, `noodlelab.verify`, for agents that write scripts,
  notebooks and tests. [Quickstart](quickstart.md) has a verified calculation in
  ten lines.
- **The Agent panel** in the editor, which runs Claude Code or Codex CLI in a
  terminal under the canvas, already connected to the MCP server.

Whichever it uses, the agent is done when `noodlelab verify --json` exits 0.

To set up a project for agents, run `noodlelab init-agent`: it writes the
instructions, a skill and `.mcp.json`. [Adopting noodlelab in a
project](adopt-in-existing-project.md) is the step-by-step, and [the guide for
agents](guide.md) is what the agent reads first.