Amaco is the open-source orchestrator that turns one chat-with-an-LLM into a supervised multi-agent flow — planning, arbitration, execution, review, verification — across Claude Code, Codex, Gemini, and any local model you can run.
One model in one chat is great for sketches. Real work — refactors, migrations, end-to-end features — needs multiple models, clear phases, and a human in the loop. Today, you stitch that by hand: tab-hopping between Claude Code, Codex, and a terminal; copy-pasting plans into chats; eyeballing diffs; losing track of cost.
Switch between Claude for plan, Codex for review, a local model for the boring bits — context evaporates with every paste.
When something works, you can't reproduce it. When it doesn't, you can't tell which model made which call.
Three vendors, three billing pages, one cursed spreadsheet. By Friday you have no idea what the experiment cost.
Write the task in markdown. Goals, constraints, files. Brief lives with the run forever.
Pick a guide — or design your own. Each guide is a graph of phases, with explicit human checkpoints.
Assign one model per role. Planner could be Sonnet; arbiter Gemini; verifier a local Llama.
A sandboxed branch opens. Phases stream. You approve where it matters. Amaco closes the loop.
Your code never leaves your machine. The orchestrator runs as a daemon, models talk to it over a local socket.
Every phase, every prompt, every gate is a YAML file you can fork. No magic, no SaaS.
The same engine drives a polished mission control, a TUI for the terminal-bound, and a scriptable CLI for CI.
MIT licensed. The CLI runs offline. The UI is a tiny local web app. Audit, fork, modify.
Guides ship with sensible defaults — but every phase is a single YAML entry. Add a "design-review" gate. Run two arbiters. Skip planning for one-liners. Fork the one closest to your team and call it your own.
Two models propose in parallel; a third arbitrates; a reviewer signs off. The safest path for high-stakes refactors.
Single planner → executor → light validation. Perfect for small, well-scoped tasks that need to land before lunch.
Architect first, then chunked execution with strict review loops between every module.
Repro → hypothesis → patch → verify, with auto re-runs until the failing test goes green. Or you stop it.
Amaco wraps the CLIs you already trust — Claude Code, Codex, Gemini — and binds local servers like Ollama and llama.cpp. No vendor lock-in, no proxy layer eating your tokens. You log in once, per vendor, with your own key.
Amaco's session ledger is a single source of truth: tokens in, tokens out, dollars spent, time taken, who did what. Slice by guide, by role, by branch. Replay any run, frame by frame. Export to a CSV your CFO will not yell about.
The UI is a thin client on top of the daemon. Anything you can do clicking around, you can do in the terminal — and pipe through CI. The daemon emits JSON; the UI is just one of many possible front-ends.
Install Amaco, point it at the CLIs you already have, and run your first orchestrated flow in under five minutes. No accounts, no proxies, no surprises.