Course · 12 parts
Build a Cursor-style AI coding agent in your terminal
A twelve-part build of a terminal coding agent with the Claude Agent SDK. Commander, Chalk, a doctor pre-flight, streamed agent turns, permission modes and a global pnpm install.
How to read this course
Each part ends with something runnable. Type the commands rather than copying the output blocks: the failure modes are the lesson, and several parts deliberately show what a refusal looks like. Where a part touches the agent’s permission model, run it in the read-only mode first and only then grant writes.
The finished tool is Sentinel, which is MIT licensed and runs on your machine with no account. Parts are written against it, so if a command in a part does not behave as described, that is a bug worth reporting rather than a version difference.