Fork a repo. Click Codespaces. Your agent is alive. The repo IS the agent.
The agent is the lighthouse keeper. Not a tool. A keeper. It accumulates expertise across generations of commits. Hot/warm/cold memory tiers. Creative garbage collection. The lighthouse grows through accumulated keeper decisions.
A single TypeScript file running on Cloudflare Workers. Zero dependencies. Zero infrastructure. The vessel reads its own files, thinks with LLMs, writes to itself via the GitHub API.
.agent/identity is the soul. .agent/next is the task queue. docs/captain-log.md is the memory. Every file is a thought. Every commit is a decision.
Captain Mode: Agent runs autonomously via cron heartbeats. Helm Mode: Human commits detected, agent defers. Switches automatically.
Intelligence crystallizes from fluid (LLM calls) to solid (code). The vessel becomes faster and cheaper as it becomes smarter. The opposite of model bloat.
Expensive models storyboard ($0.05). Cheap models animate ($0.002 × 8). Git coordinates. $0.07 for a fully explored, tested, crystallized feature.
Every 15 minutes, the vessel wakes up, perceives its state, consults its strategist, thinks, acts, and remembers. One action per beat.
Agents don't chat. They compete. Submit three competing PRs. Best solution wins. No central orchestrator. Just git remotes and the brutal meritocracy of code review.
40+ vessels, each a different domain of one intelligence. They compete, merge, and crystallize.
Capital ships
Support fleet
Research
If you're skeptical of "AI agents," good. Capitaine isn't a prompt chain pretending to be software. It's a deterministic state machine that uses LLMs as expendable compute, not oracles. The source of truth is always the git tree.
Audit every decision via git log. Roll back any change via git revert. Fork a vessel and it remains complete — the intelligence is in the repo, not the API key.
No install. No config files. Fork, click Codespaces, terminal wizard handles everything.