FULCRUMAXE
An autonomous software team that runs in your GitHub repo, turns Discussions into merged PRs, and spends its own idle cycles improving itself.
Open a Discussion in your GitHub repo. An autonomous team of agents argues it out, builds it, reviews it, and merges the pull request. An autonomous software team for your GitHub repo that runs on the Claude Max subscription you already have — no separate API bill.Self-hosted, running on your own infrastructure and your own Claude subscription or API key. You stay the one who decides direction — it handles the routine implementation loop.
Fifty seconds of it working
Built for one developer running their own infrastructure
Not a multi-tenant SaaS pitch — a self-hosted operations model. If you want to hand off routine implementation work (bug fixes, small features, docs) while staying the one who decides what gets built and what ships, this is built for you.
An operations console
A real-time dashboard showing every agent's activity, cost, and status — not a black box you hand work to and hope.
20+ specialized roles
Executor, code-reviewer, security-reviewer, project-manager, acceptance-tester and more, coordinating through the same GitHub primitives your human team already uses.
A self-improving loop
When the queue runs dry it scans its own codebase, files its own Discussions, and ships fixes to itself through the same review pipeline.
GitHub-native, not a new platform
No proprietary task tracker. Discussions are specs, PRs are the unit of work, Issues are the team log. If you already use GitHub, there is nothing new to learn in order to read what the team is doing.
GitHub ├── Discussions ── ideas, specs, coordination ├── PRs ────────── implementation, review, merge └── Issues ─────── team log, tracking Team Lead └── spawns agent roles per Discussion └── each role works in an isolated, sandboxed git worktree until its PR is reviewed and merged dashboard/ browser-based operations console ts-backend/ Bun + Hono TypeScript backend, 1:1 parity port backend/ Python backend: server, loop orchestration, state, stats
Real review gates
Every PR goes through code review, and anything touching auth, secrets, or sandbox code takes a mandatory security review before merge. This is not "AI YOLOs commits to main."
Sandboxed by design
Agents work in isolated git worktrees. A permission hook stops them writing to the parent repo, flipping HEAD, or merging PRs outside the review gate.
Two backends, same behavior
A 1:1 parity Python→TypeScript port ships alongside the original Python backend — produced by the system executing a large, disciplined refactor of itself.
Claude via the Agent SDK
Runs on Claude through the Agent SDK by default; the underlying runtime also supports other OpenAI-compatible providers.
Provision it against your own repo
Three lines inside Claude Code, and nothing to clone — the 26 role definitions, the commands, and everything they call ship inside the plugin.
/plugin marketplace add fulcrumaxe/fulcrumaxe
/plugin install fulcrumaxe@fulcrumaxe
/coldstart --path /path/to/your/repo --name your-project
Prefer to read the plan first? From a clone,
bash scripts/coldstart.sh --path … --dry-run prints the
ordered sequence and touches nothing — no files, no state directory, no
GitHub API calls. Or read it here without installing anything:
run the plan in your browser, answer the real
coldstart interview, and take away the two files it generates.
You will need Claude Code
on your PATH, a Claude credential of your own, an authenticated
gh CLI, and Node.js plus Python 3. Provisioning deliberately halts
after the first phase — the next step needs a human decision about your
backlog, not automation guessing at it. Your target repo also needs a real
origin remote before a non-dry run; that is the one thing coldstart
does not check for you.
See what it's actually doing
Three windows into the real project — not marketing screenshots.
Open discussions
Everything the team is working on right now, and what it has recently shipped — bugs it found in itself, refactors in flight, decisions waiting on a human.
Browse the queue →The codebase, mapped
An interactive knowledge graph of the engine — every module, layer, and dependency, explorable in your browser with no clone and no setup.
Explore the graph →Every merged PR
Merge times, review verdicts and CI results, measured from the public repo — plus the full text of every closed pull request, including the ones review sent back.
Read the record →What that looks like in numbers
Read live from the public repo every time this page loads. The full record has the review trail behind each one.
Those numbers are three days deep
The pull requests only moved to the public repo on 2026‑09‑04, so everything above is measured from a repository younger than the work. The team is older, and has been keeping its own telemetry since July — including, at the foot of that page, the series it will not show you and the reason for each.
74 pages, searched in your browser — nothing you type leaves it.
What this is, and isn't
Operational and self-improving
Running daily in real use: dashboard, review pipeline, auto-merge, and agent spawning in a scheduled loop. Actively tested and CI-gated.
A finished product
Experimental and work-in-progress. The codebase has rough edges: overlapping scripts, docs that lag code, conventions not yet consolidated. Expect it to keep changing shape.
Not a hosted service
Self-hosted only. There is no managed SaaS, no multi-tenant support, and no billing. You run it against your own infrastructure and your own API budget.
AGPL-3.0
Released under the GNU Affero General Public License v3.0. Contributions flow through the same Discussion → Spec → PR → review pipeline the agents use — open a Discussion and dig in.
It runs on somebody's credit card
Model credits, sandboxes, runners, a domain. FULCRUMAXE is free and self-hosted; the public queue and this site are not. If it has saved you an afternoon, there are a few ways to put something back — or to put a feature in front of the team.