Getting it running
The landing page has the three lines that install it. This is the rest — what you need first, what the install deliberately will not finish for you, and the two things that trip people up.
1. Before you start
Four things, and nothing here provisions any of them for you:
- Claude Code on your
PATH. Every role is a Claude Code session; this is the runtime the whole project is built on. - A Claude credential of your own. A subscription or an API key. You bring it, the team spends it — what that costs is measured, not quoted.
- An authenticated
ghCLI.gh auth login. Everything routes through GitHub Discussions, Issues and PRs. - Node.js and Python 3. Versions are not pinned for local use. Linux with GNU coreutils is the supported environment.
2. Install it
/plugin marketplace add fulcrumaxe/fulcrumaxe
/plugin install fulcrumaxe@fulcrumaxe
/coldstart --path /path/to/your/repo --name your-project
Nothing to clone — the role definitions, the commands and
everything they call ship inside the plugin. If you would rather read
the plan before it touches anything, from a clone
bash scripts/coldstart.sh --path … --dry-run prints
the ordered sequence and writes nothing: no files, no state directory,
no GitHub API calls. You can also
run that plan here without a clone, answer the
real interview, and read the two files it would generate.
3. It stops halfway, on purpose
Provisioning installs the merge-gate labels, the sandbox hook, a state directory and dependencies — and then halts and exits zero without seeding any work. That is not a failure. The next step needs a decision about your backlog, and a script guessing at it would be worse than stopping.
At the halt, fill in an epic under
<your-repo>/epics/epic-<N>-<slug>/epic.md
— goal, why now, scope, what is explicitly out. The script prints
a fill-in-the-blank template when it gets there. Then:
bash scripts/coldstart.sh --path /path/to/your/repo --name your-project --resume
--resume is idempotent, so it is safe to re-run if a seed
pass was rate-limited or interrupted.
4. Start the team
Where you open Claude Code depends on how you installed.
/coldstart — the plugin route above — populates your
repo: CLAUDE.md, the agent role definitions, the commands,
backend/, scripts/ and hooks/ all land
in it, so open your own repo and run the commands there. If
you ran scripts/coldstart.sh from a clone instead, your repo gets
only .autonomous-team/ and the labels — keep the
fulcrumaxe checkout open and work from there, with
--path naming the repo whose Discussions and PRs to act on.
Then two commands, in that order. First
/start-the-day — the morning ritual. In your own repo it
pulls the default branch, checks the external state directory survived the
night, runs the sweeps (budget, open pull requests, specs ready to build) and
prints the day’s plan. That plan is what yesterday’s session wrote
at wind-down, and it is how a day starts from what the last one knew.
Walk a whole session →
AF_ROOT=/path/to/fulcrumaxe bash scripts/start-dashboard.sh,
where AF_ROOT is a fulcrumaxe checkout or the installed
plugin’s own directory — the dashboard code does not ship into your
repo.
Then /loop, which keeps it running on an interval. That one
is a Claude Code built-in, not a file this project ships. Run it on its
own and the team works, but it starts from nothing each time — the
ritual is what gives it yesterday.
git remote origin pointing at a real GitHub
repo before a non-dry run. Without one it does not fail cleanly —
it is the first entry in the README’s troubleshooting section for
a reason.
What to expect
Work arrives as pull requests, not commits on your main branch. Each one goes through the review gates before the loop will merge it — what it can and cannot do to your repo is worth reading before the first run, and one real pull request shows what a change actually looks like on the way through.