Your backlog,
done by breakfast.
Breakfast is a macOS app that runs Claude Code on your task queue while you're away. Every session works in its own isolated Git worktree, commits what it finishes, and reports honestly — done, blocked, or paused — by morning.
macOS 26+ · Runs on your existing Claude Code subscription
One agent session is manageable.
Real work is chaos.
Multiple projects, branches, usage windows, task state — you either babysit everything or trust too much autonomy and wake up to a mess with no record of what happened. Breakfast is the middle: not babysitting, not blind autonomy — controlled, orchestrated work.
Humans steer. Agents execute.
Breakfast is a native macOS orchestration layer for autonomous Claude Code sessions. It turns scattered terminal work into organized, context-aware workflows: queue tasks across projects, carry context forward from finished work, run each task through a multi-step loop, pace against usage limits and timing, and keep every execution reviewable in its own local worktree. Instead of babysitting terminals or handing everything to blind autonomy, Breakfast keeps you in control while your sessions keep the work moving.
Queue it tonight.
Review it tomorrow.
- 01
Queue your tasks
Point Breakfast at your projects. Tasks are plain Markdown backlog files that live in your repo, versioned with your code.
- 02
Start a session
Breakfast launches Claude Code in a terminal it orchestrates — one isolated Git worktree per run. Main stays untouched.
- 03
The loop runs
Claude works task by task, up to 50 iterations per session. Breakfast tracks phases, liveness, and progress, and commits finished work against its task ID.
- 04
Wake up to results
Completed work lands on main — or opens a pull request, your choice per project. Blocked work is reverted, labeled, and waiting for your call.
One task, one worktree,
refined until it's right.
Every task spins up its own loop — planning, execution, and verification, iteration after iteration, up to the merge or PR at the end. Those iterations aren't brute force: Breakfast watches each one — scoping the work, catching errors, pacing against usage limits, and flagging when a task genuinely needs a human instead of grinding on. It keeps each branch clean and lands work only once it's truly verified, and you run the whole loop system from a GUI, not a wall of terminals.
Autonomy you can leave alone.
Isolated worktrees
Every session runs on its own branch in its own worktree. Nothing touches main until the work is verified.
A board you can steer
Breakfast keeps a Kanban board of every task: the AI moves cards as work runs, and you can read, reorder, and edit it any time. Decide what runs next across projects before pressing play, with a live phase timeline for each session.
Context that compounds
Tasks and their implementation notes live as Markdown in your repo. Finished work stays on the record, so every new session reads the history and builds on it instead of starting cold.
Multi-project orchestration
Round-robin across projects with per-project concurrency caps and session timeouts.
Scriptable, CLI-driven
Everything the app does, the breakfast CLI does too — status, queue, start and stop, worktrees, history. Wire it into your own scripts, or let agents like OpenClaw and Hermes drive Breakfast for you.
Honest outcomes
Sessions end in a typed result: completed, blocked, or usage limit. The agent can raise its hand instead of pretending.
Runs on your subscription
Breakfast drives your real Claude Code sessions in a terminal on your Mac, signed in with your existing login — no API wrapper, no hosted orchestrator, no per-token bill from us.
Usage-aware pacing
When Claude hits its usage limit, Breakfast stops cleanly and resumes on your next session instead of burning retries.
Manageable by design.
Every session ends in a typed outcome — so you steer the work and review real results, instead of guessing what happened overnight.
- Completed
All tasks done, tests green, work committed to your branch. - Blocked
Something genuinely in the way — reverted, labeled, waiting for you. - Usage limit
Claude's limit hit — paused cleanly, resumes next session.
Manual terminals
One prompt at a time, watched
You notice failures by luck
Manual usage-limit tracking
You babysit
- Recommended
Breakfast
Queued, orchestrated sessions
Typed outcomes: completed, blocked, usage limit
Usage-aware pacing, clean pause + resume
You orchestrate
Blind autonomy
Fire-and-forget, unattended
Failures hide until you look
Unpredictable, wasted retries
You lose control
Breakfast is in beta.
It runs on real backlogs today, but it isn't on sale yet — pricing and downloads are on the way.
Questions, answered.
What do I need to run Breakfast?
A Mac running macOS 26 or later, Git, and a Claude subscription that includes Claude Code. Breakfast bundles the Claude Code CLI and signs in with your existing account.
What does an autonomous session actually do?
Breakfast picks the next task from your queue, creates a fresh Git worktree, and runs Claude Code in a loop until the work is done or it genuinely can't continue. Finished work is committed with its task ID and lands on main (or a PR) only after the session verifies it. You get a full session log either way.
Does Breakfast use the Claude API or my subscription?
Your subscription. Breakfast runs your real Claude Code sessions in a terminal it orchestrates on your Mac — signed in with your existing Claude Code login, not routed through an API wrapper or a hosted service. There's no server in the middle, so your subscription works exactly as it does in your own terminal.
Is my code private?
Breakfast runs entirely on your Mac and has no server component reading your repository. Your code goes only where Claude Code already sends it — to Anthropic, under your own account and their terms.
What will it cost?
Pricing isn't final yet, so we're not going to quote you a number we might change. Breakfast is in beta and not on sale — we'll announce pricing here before launch.
Does it work with tools other than Claude Code?
Not yet. v1 is built for Claude Code only. If Breakfast is useful, more agents may follow.
What happens when Claude hits a usage limit?
Breakfast detects it, ends the session cleanly, and marks it as paused — no wasted retries. The next session picks up where it left off.