Run Claude Code overnight: a fleet that works your board while you sleep

What goes wrong when you leave agents running

Starting five Claude Code sessions before bed is easy. Waking up to something useful is not. The usual failures:

• Two agents take the same task, and you get two branches fixing one bug.
• An agent picks up something that was not ready: half-specified, blocked by other work, or needing a decision only you can make.
• An agent stops at the first question and waits all night for an answer.
• An agent asks permission for a command and sits on the prompt for eight hours.
• Nobody checks the work before you look at it, so the morning starts with a review queue.
• Nothing records what happened, so you read transcripts to find out.

Each one has a fix. Put together, they are a fleet.

The four parts of an overnight run

1. A queue of ready work

An agent should only take a task that is ready: specified, unblocked, unclaimed and not needing a person. That means a board that knows stages, blockers and claims. In Projectify, Start next already answers "what is the highest-ranked task an agent can take now", and the fleet takes exactly what Start next would hand an agent. Epics are never taken; tasks that need a person, a browser or design work are skipped. Any single task can be kept off the fleet.

2. One isolated session per task

Each task runs as a background Claude Code session in its own git worktree:

claude --bg --worktree hbr-47 --agent projectify:projectify-worker

You do not type this; the launcher does. Each worker claims its task with its worktree and branch, so no other session takes it, then follows the project's workflow, pushes its branch and opens a pull request with gh. It moves the task to Review with the PR and ends with a handoff. It never merges, and a push to main asks first.

3. Someone to answer while you sleep

Workers ask questions. Overnight, the Coordinator answers them. It is a Claude Code session that drains an inbox: questions a worker asked, approvals, comments on tasks the fleet is working, and room posts that mention @fleet. You decide per project, per category, what it may do:

Decide answers and the worker carries on. Recommend records an answer for you to confirm. Ask me leaves it in Needs you. You can overrule any answer by answering again. The Coordinator never edits code.

4. A review before you look

When a pull request opens, a Reviewer reads it and runs the repository's checks. Its verdict is ready or changes; changes start a developer rework on the same branch, up to the number of review rounds you set. By morning, what is in Review has already been through one pair of eyes.

Set it up

You need Claude Code, Node 18 or later, and gh signed in.

1. Install the plugin in Claude Code:

/plugin marketplace add isurfraz/projectify-plugin
/plugin install projectify@projectify

1. Set up the machine: run /projectify:setup in any repo and say yes to the fleet. It finds your repos, trusts each folder for Claude Code and ends with projectify fleet doctor.
2. Switch projects on in the web app: People › Fleet, or the project's Setup › Fleet for limits, triggers, models and the Coordinator's policy.
3. Run it:

projectify fleet            # every switched-on project with a repo here
projectify fleet HBR WEB    # only these projects
projectify fleet --dry-run  # list what it would take, then exit

The launcher waits on the board over HTTPS, so waiting costs no tokens. It prints the projects it works and skips, with why, then one line per event.

Models and effort per role

Planner, developer, reviewer and coordinator can each have their own model and effort. Set your own defaults in People › Fleet › My defaults and override them per project. Developer rules pick a model by task size or kind: a quick model for small fixes, a stronger one for large tasks. A fallback model takes over when the first is overloaded.

Keep watching, or stop when done

Each machine either keeps watching for new ready work or stops when nothing is ready, running or waiting in the inbox. projectify fleet --watch or --until-empty overrides it for one run. Watching starts no Claude session until something happens.

In the morning

• Review holds the pull requests, each with its task, its handoff and the Reviewer's verdict.
• Needs you holds what the Coordinator left for you: risky approvals, product calls, recommendations to confirm.
• Activity shows every change, attributed: "Claude for Isaac", and which role did it.
• projectify fleet status or the live fleet panel shows what is still running. Ctrl-C once stops after the current tasks, twice stops now.

Limits worth knowing

• The fleet is for Claude Code in a terminal. On claude.ai and Claude Desktop you work chat by chat.
• Workers run on your machine and your Claude plan; the machine has to stay awake.
• Each machine has a worker cap and each project its own limits. Start with two or three workers and raise it once the reviews look good.
• Overnight work is only as good as the tasks. Write a clear summary and acceptance for each, and keep anything vague in the Define stages until it is ready.

Cost

The fleet is included on every Projectify plan. Solo is free for one person, with every feature. Team is $10 per user per month billed yearly or $12 monthly, and agents never take a seat: a fleet of five workers costs the same as none.