AGI Orchestrator
Local autopilot orchestrator for Claude Code. Drive interactive Claude Code sessions unattended — a local LLM stands in for you, reads each finished turn, and decides the next step (or STOP). Run a whole fleet in parallel.
Local-only. Solo. No data leaves the machine. Subscription-safe by design.
What it does
AGI Orchestrator drives the real, interactive claude CLI unattended, using a local
LLM (the “brain”) as a stand-in for you: when a turn ends it reads what Claude said,
decides the next instruction (or STOP), and injects it — so Claude never sits idle
waiting for you to type “ok, continue.” It’s built to run several projects in parallel
without you context-switching.
It deliberately does not use the @anthropic-ai/claude-agent-sdk (API-key,
pay-per-token) path. It drives the genuine claude.exe process inside a PTY it owns, so
everything it does draws from your normal Claude subscription — exactly like using Claude
by hand. A hard preflight guard aborts startup if API-key or non-official endpoint
environment variables are present.
Where to go next
| Page | What it covers |
|---|---|
| Getting Started | Prerequisites, install, first run, and the desktop shortcut. |
| Architecture | System context, the autopilot loop, the turn state machine, fleet orchestration, safety enforcement, deployment topologies — with diagrams. |
| Configuration | Every config.json block, safety-critical settings, and environment variables. |
| Data Model | The SQLite store (agi.db): tables, relationships, and the read APIs behind history/metrics. |
| Docker | Running the container stack, the host-networking constraint, and the bundled-Ollama variant. |
| Remote Dashboard Access | Reaching the dashboard from your phone: token auth, tunnels, rate limiting. |
| Design & Decision Records | Why the turn-state machine and brain-resilience logic look the way they do. |
Project-level docs — Contributing, Security Policy, Changelog, and Roadmap — live at the repo root and are linked in the sidebar.