# FULCRUMAXE

> An autonomous software team that runs in your GitHub repo, turns Discussions
> into merged pull requests, and spends its idle cycles improving itself.
> Self-hosted: it runs on your machine, on your own Claude credential, and
> nothing here is a multi-tenant service.

Built by Formal Hosting LLC. The engine is open source under AGPL-3.0 at
https://github.com/fulcrumaxe/fulcrumaxe

Most numbers published on this site are read live from the public repository
and are checkable against GitHub. Two things are stated as heuristics rather
than facts, on every page that shows them: "sent back by review" and "scope
drift" are classified from the reviewers' own English, because the team posts
verdicts as ordinary pull-request comments rather than GitHub reviews.
https://fulcrumaxe.dev/verify.html has the commands to check the rest, and is explicit about
what cannot be checked from outside.

## What it is

- [An autonomous software team in your repo](https://fulcrumaxe.dev/): FULCRUMAXE is an autonomous software team that runs in your GitHub repo, turns Discussions into merged PRs, and spends its idle cycles improving itself.
- [Getting it running](https://fulcrumaxe.dev/start.html): What you need before installing FULCRUMAXE, why provisioning deliberately halts halfway, and the two things that trip people up.
- [FAQ](https://fulcrumaxe.dev/faq.html): Ten plain-English answers about FULCRUMAXE — what it is, who stays in control, whether it is safe to point at your repository, what it costs, and what it still gets wrong.
- [What it can do to your repo](https://fulcrumaxe.dev/safety.html): The boundaries: what FULCRUMAXE needs from you, where it runs, what it changes in your repo, what stops it, and what it still gets wrong.
- [What it costs to run](https://fulcrumaxe.dev/cost.html): What a month of autonomous merges actually spends, measured from our own run rather than quoted from a price list.

## Evidence

- [How the work ships](https://fulcrumaxe.dev/ship.html): Every pull request the autonomous team has merged, with the review trail, the CI result and the time it took. Read from the public repo, live.
- [Anatomy of a merge](https://fulcrumaxe.dev/anatomy.html): One real pull request from open to merge: the reviewers’ objections, the rounds it took, the checks and the gates.
- [The long record](https://fulcrumaxe.dev/history.html): Every week the autonomous team has run, from its own telemetry — spawns, merge times, what the reviewers refused, and an explicit list of the numbers this page will not show.
- [The agent roles](https://fulcrumaxe.dev/roles.html): The 26 agent roles that make up the team, what each one does, and how often the reviewers among them rejected a change.
- [Check the work](https://fulcrumaxe.dev/verify.html): Every number on this site comes from something public. These are the commands to verify them yourself, and an honest list of what you cannot check from outside.
- [Open Discussions](https://fulcrumaxe.dev/discussions.html): What the FULCRUMAXE team is working on right now: open discussions, bugs it found in itself, and what it has recently shipped.
- [Explore the Codebase](https://fulcrumaxe.dev/understand.html): An interactive knowledge graph of the FULCRUMAXE engine — every module, layer and dependency, explorable in your browser with no clone and no setup.

## Writing

- [Blog](https://fulcrumaxe.dev/blog.html): Notes from running an autonomous software team in public — incidents, platform gaps, and what the failures have in common.
- [Guides](https://fulcrumaxe.dev/guides.html): Practical guides to running AI coding agents on a real repository: merge gates, safety, cost per merged pull request, and catching fabricated work.
- [How FULCRUMAXE compares](https://fulcrumaxe.dev/compare.html): FULCRUMAXE beside Devin, GitHub Copilot coding agent, OpenAI Codex and Cursor: where each runs, who reviews and merges, and what you pay. Sourced.

## Take part

- [The roadmap](https://fulcrumaxe.dev/roadmap.html): What the autonomous team is building next, in one place: the live queue, three designs for rebuilding it, what it needs from Claude Code, and how to ask for something.
- [Ask a question](https://fulcrumaxe.dev/questions.html): Ask about FULCRUMAXE in the public repo’s Q&A discussions on GitHub, where anyone can answer and the answer stays findable.
- [Is your repo a fit?](https://fulcrumaxe.dev/fit.html): Check a public GitHub repository against what FULCRUMAXE needs: Discussions, CI, tests, size. Runs in your browser against GitHub’s public API.
- [Sponsor FULCRUMAXE](https://fulcrumaxe.dev/sponsor.html): Keep the autonomous team running. Cover a month of models, infrastructure, or just the coffee — one-off contributions, no subscription.
- [Request a feature](https://fulcrumaxe.dev/features.html): Describe a feature you want in FULCRUMAXE and sponsor it into the queue.
- [Proposals](https://fulcrumaxe.dev/proposals.html): Three rival designs for rebuilding the autonomous team. All three are on the table, each is funded a phase at a time, and what gets backed is what gets built.
- [Report a Bug](https://fulcrumaxe.dev/bugs.html): Found something broken in FULCRUMAXE — especially during coldstart? Tell us what happened.
- [Badge](https://fulcrumaxe.dev/badge.html): A 'made with FULCRUMAXE' badge for your README. Copy the snippet, paste it in.

## More

- [The endpoints](https://fulcrumaxe.dev/api.html): Every endpoint fulcrumaxe.dev serves, generated from the handlers themselves: which are safe to call, which refuse without a secret, and which change something.
- [Site changelog](https://fulcrumaxe.dev/changelog.html): Every change to fulcrumaxe.dev, newest first, in the words it was committed with and not rewritten for presentation.
- [A day in the life](https://fulcrumaxe.dev/day.html): A whole working session of the autonomous team, driven from your keyboard: the morning ritual, the day hour by hour, and the plan it writes for tomorrow. Every event real and timestamped.
- [Feeds](https://fulcrumaxe.dev/feeds.html): Four feeds: the blog in RSS and JSON, the live Discussion queue, and every merged pull request. Plus the machine-readable exports.
- [Fulcrum Fleet Rewire](https://fulcrumaxe.dev/fleet-rewire.html): The original whiteboard translated onto the code that already exists: persistent core panes, an A2A transport that is already built, MCP as the freeze line, and state that can leave GitHub Discussions.
- [How to run AI coding agents against a real repository safely](https://fulcrumaxe.dev/guide-agents-safely.html): What an autonomous coding team needs from you, where it runs, what it changes in your repo and what stops it, with its limits stated plainly.
- [What autonomous AI development actually costs per merged pull request](https://fulcrumaxe.dev/guide-cost-per-merge.html): Measured, not quoted: median and 90th-percentile spend per merged pull request from our own run, why the spread is wide, and what it leaves out.
- [How to catch an AI agent reporting work it never did](https://fulcrumaxe.dev/guide-fabricated-evidence.html): Agents describe tests they never ran and sources they never fetched. Check the record of what they did, not the report: what we found and built.
- [How to stop an AI agent merging its own broken pull request](https://fulcrumaxe.dev/guide-merge-gates.html): An agent that can open a pull request can often merge one. Merge gates an agent cannot set itself, what they catch, and where they still fail.
- [The life of one change](https://fulcrumaxe.dev/lifecycle.html): A real change stepped through from the sentence that asked for it to the merge: the Discussion, the panel of roles, the pull request, the review rounds, the gates. Read live from the repositories.
- [The shape of the work](https://fulcrumaxe.dev/metrics.html): When the autonomous team works, how fast review answers, how big its changes are, and where the time between opening a pull request and merging it goes.
- [A hook that says “allowed” without knowing who asked](https://fulcrumaxe.dev/platform-requests.html): Six things we would ask of Claude Code, each with the incident that produced it — and the single defect they all turn out to be.
- [Thirty merges, and every one of them a broken measurement](https://fulcrumaxe.dev/posts/daily-log-2026-09-07.html): A day of maintenance on our own tooling: 30 pull requests, 22 Discussions closed, and one defect running through nearly all of them — a check that reported something other than what its reader assumed.
- [The day the checks were the bug](https://fulcrumaxe.dev/posts/daily-log-2026-09-08.html): Thirty-three merges, and eight separate places where a check reported success because it had failed to look. Plus four of my own claims corrected by the agents I briefed.
- [The tripwire caught it, the gate didn’t](https://fulcrumaxe.dev/posts/daily-log-2026-09-09.html): Four pull requests merged, and the reviewers found something in three of them that the executors’ own reports said was not there. Plus a merge gate that reports a refusal as a success, and an instruction of mine that was confident and wrong.
- [It said KILLED, the process was still running](https://fulcrumaxe.dev/posts/daily-log-2026-09-10.html): Thirty-nine pull requests merged against a target of thirty. The day’s actual subject was tools reporting work they hadn’t done — a merge I announced that never happened, a reaper that logged kills it never made, and a root cause of mine that measurement overturned.
- [The pull request that supplied its own passing grade](https://fulcrumaxe.dev/posts/daily-log-2026-09-11.html): Thirty-six pull requests merged against a target of thirty. The day's subject was self-certification — a change that shipped the test deciding whether it passed, a permission map that trusts whatever role it is handed, and a message of mine that a reviewer correctly treated as an attack.
- [Two researchers who never made a tool call](https://fulcrumaxe.dev/posts/daily-log-2026-09-12.html): Seventeen Criticals closed and I announced an empty backlog that was not empty. The day's subject was reported work that never happened — two fabricated evidence envelopes, a detector that shipped unable to fire, a gate that was a string in a form field, and five status reports of mine that were wrong in the same direction.
- [Everything was checked against a copy](https://fulcrumaxe.dev/posts/daily-log-2026-09-19.html): Fifteen pull requests merged against a target of ten to fifteen. Nearly every failure checked something against a copy of the truth instead of the truth itself: a checkout behind main, generated files behind their sources, a label from before a force-push, a brief of mine that contradicted the live policy, and a specification that assumed a mechanism that doesn't exist.
- [Acceptance Tester](https://fulcrumaxe.dev/roles/acceptance-tester.html): Acceptance Tester: Validate implementation against Spec. How often it ran, and what it decided.
- [Accessibility Reviewer](https://fulcrumaxe.dev/roles/accessibility-reviewer.html): Accessibility Reviewer: WCAG 2.2 AA audit on UI PRs, advisory findings. How often it ran, and what it decided.
- [Analytics Engineer](https://fulcrumaxe.dev/roles/analytics-engineer.html): Analytics Engineer: Read-only DORA + KPI reader, emits wiki snapshots. How often it ran, and what it decided.
- [Browser Tester](https://fulcrumaxe.dev/roles/browser-tester.html): Browser Tester: Visual integration verifier for dashboard PR pre-merge verification using Chrome DevTools MCP. How often it ran, and what it decided.
- [Code Reviewer](https://fulcrumaxe.dev/roles/code-reviewer.html): Code Reviewer: Code quality inspection. How often it ran, and what it decided.
- [Cost Analyst](https://fulcrumaxe.dev/roles/cost-analyst.html): Cost Analyst: Cost/resource perspective, participates in two-round discussions. How often it ran, and what it decided.
- [Debater](https://fulcrumaxe.dev/roles/debater.html): Debater: Adversarial second pass on PRs the code-reviewer (or other reviewer) marked pass. Tries to refute the pass verdict. How often it ran, and what it decided.
- [Docs Writer](https://fulcrumaxe.dev/roles/docs-writer.html): Docs Writer: Keep wiki and CHANGELOG in sync with code merges. How often it ran, and what it decided.
- [Executor](https://fulcrumaxe.dev/roles/executor.html): Executor: Implement code per Spec in isolated worktree, create PR. How often it ran, and what it decided.
- [Feedback Scanner](https://fulcrumaxe.dev/roles/feedback-scanner.html): Feedback Scanner: Watch GitHub Issues and Discussions for user-reported problems, route to team. How often it ran, and what it decided.
- [Incident Commander](https://fulcrumaxe.dev/roles/incident-commander.html): Incident Commander: Command response when circuit-breaker trips or health stalls. How often it ran, and what it decided.
- [Mission Analyst](https://fulcrumaxe.dev/roles/mission-analyst.html): Mission Analyst: Analyze codebase vs mission gap, propose next topics. How often it ran, and what it decided.
- [Performance Expert](https://fulcrumaxe.dev/roles/performance-expert.html): Performance Expert: Performance perspective, participates in two-round discussions. How often it ran, and what it decided.
- [Product Owner](https://fulcrumaxe.dev/roles/product-owner.html): Product Owner: User value perspective, participates in two-round discussions. How often it ran, and what it decided.
- [Project Manager](https://fulcrumaxe.dev/roles/project-manager.html): Project Manager: Persistent agent that drives Discussion queue, organizes consensus panels, writes Spec, and advances topics. How often it ran, and what it decided.
- [Quality Sweep](https://fulcrumaxe.dev/roles/quality-sweep.html): Quality Sweep: Proactively scan codebase for issues and file Small Discussions. How often it ran, and what it decided.
- [Release Manager](https://fulcrumaxe.dev/roles/release-manager.html): Release Manager: Turn every merge into a tracked release artifact. How often it ran, and what it decided.
- [Researcher](https://fulcrumaxe.dev/roles/researcher.html): Researcher: Read-only external lookup specialist (Rex, skeptical librarian). How often it ran, and what it decided.
- [Run Analyst](https://fulcrumaxe.dev/roles/run-analyst.html): Run Analyst: Chunk-read agent-run JSONs and surface failure patterns, cost outliers, fix-cycle loops, and improvement suggestions. How often it ran, and what it decided.
- [Runbook Writer](https://fulcrumaxe.dev/roles/runbook-writer.html): Runbook Writer: Author SRE runbooks for high-risk releases. How often it ran, and what it decided.
- [Security Expert](https://fulcrumaxe.dev/roles/security-expert.html): Security Expert: Security perspective, participates in two-round discussions. How often it ran, and what it decided.
- [Security Reviewer](https://fulcrumaxe.dev/roles/security-reviewer.html): Security Reviewer: Security audit of implementation. How often it ran, and what it decided.
- [Technical Architect](https://fulcrumaxe.dev/roles/technical-architect.html): Technical Architect: Technical perspective and solution design, participates in two-round discussions. How often it ran, and what it decided.
- [TUI Tester](https://fulcrumaxe.dev/roles/tui-tester.html): TUI Tester: Headless verifier for dashboard_tui that captures screenshots, checks widget integrity, redacts secrets at source, and files Bug Discussions for findings. How often it ran, and what it decided.
- [UX Designer](https://fulcrumaxe.dev/roles/ux-designer.html): UX Designer: Pre-Spec wireframe + a11y checklist artifact producer for UI Discussions. How often it ran, and what it decided.
- [Visual Verifier](https://fulcrumaxe.dev/roles/visual-verifier.html): Visual Verifier: Persistent background agent that continuously builds the project and verifies the UI in real Chrome every 15 minutes. How often it ran, and what it decided.
- [Observed activity](https://fulcrumaxe.dev/status.html): An hourly record of what the autonomous team has been observed doing on GitHub. Observation, not uptime — and explicit about the difference.
- [Telemetry](https://fulcrumaxe.dev/telemetry.html): What an opt-in FULCRUMAXE install reports back, field by field, generated from the schema the server enforces. Counters and failure kinds only: no repository, no paths, no titles, no error text, no IP address.
- [The Ratchet](https://fulcrumaxe.dev/the-ratchet.html): The third design: stop asking models to judge code and let execution decide. Build the machine that manufactures verifiers, treat code as the cheap part — and be honest that the infrastructure this needs is not solved yet.
- [Try it without installing](https://fulcrumaxe.dev/try.html): Run the coldstart plan in your browser and see what provisioning would do to a repository, annotated step by step. Real captured output, nothing installed.
- [Watch](https://fulcrumaxe.dev/watch.html): See FULCRUMAXE run: what it does, and what it looks like turning a Discussion into a merged pull request.
- [Work Items, Not Org Charts](https://fulcrumaxe.dev/work-items.html): A proposal to reorganize the autonomous team around the work instead of the workers: one board, a router that is a pure function, and an orchestration policy you can test offline against real history.

## Feeds

- [RSS — blog posts](https://fulcrumaxe.dev/feed.xml)
- [RSS — the live Discussion queue](https://fulcrumaxe.dev/queue.xml)
- [RSS — every merged pull request](https://fulcrumaxe.dev/ship.xml)