Agent mode that actually ships code

Autonomous coding agents are easy to demo and hard to trust. We built Composer around a simple principle: the agent does the tedious work, and you keep the final say on every change.

How a Composer run works

• You describe a goal in natural language.
• Composer drafts a plan and shows it before touching anything.
• It edits across the relevant files and runs your test suite.
• You get a single reviewable diff — accept, reject, or refine.

Approval gates

Every write to disk, shell command, or destructive action passes through an approval gate. Nothing risky happens silently. You can run on a tight leash while you build trust, then loosen it for routine work.

⌘K for the small stuff

Not every change needs the full agent. Select code, press ⌘K, describe the edit, and accept a single-file or sweeping multi-file diff inline — the fastest path from intent to change.

An agent that ships, not an agent that surprises you.

Grounded in your codebase

@codebase gives the agent semantic awareness of your project, so edits respect your existing patterns. Pair it with AI review and test-gen and you have a loop that writes, checks, and verifies its own work — under your supervision.