Coding agents

blu doesn’t write code itself. It runs the agents you already use — Claude Code and Codex — over the Agent Client Protocol (ACP). One agent process per agent kind serves all your projects; each conversation is its own ACP session.

Setup is in Install.

Picking the agent and the folder

In the project editor:

1. Agent — Claude, Codex, or the test agent.
2. Folder — required for a real agent: the agent works in it. Without a folder the task fails at once.
3. Linked folders — more folders the agent may read and change, say the backend next to the frontend. Make it the main one swaps which folder the agent works in.

Dropping folders from Finder onto a blob does the same: the first becomes the main folder, the next ones are linked.

A git worktree for every task

With A git worktree for every task on (the default for new projects) and the main folder in a git repository, the agent works in a separate working copy on a blu/… branch:

• your own working copy stays exactly as it is;
• finished changes are committed to that branch, one commit per task;
• merge them whenever you like: git merge blu/…;
• the result panel shows the branch, Show in Finder and Remove the working copy.

A conversation keeps its worktree across tasks, so a follow-up continues where the agent left off. Removing the working copy keeps the branch if it has commits of its own.

Without git or with worktrees off, the agent works in the folder itself, and the project runs one task at a time.

Sign-in

If an agent answers that it needs a sign-in, the result panel shows Sign in in Terminal. blu opens Terminal with the agent’s login command; the tokens stay with the agent.