A local-first control plane for your coding agents

You keep using the coding agents you already trust. MUON sits above them as an orchestrator: it plans work, dispatches it to a fleet of agent instances (0–3 per vendor), watches what they do, and stops for your decision at every point that matters — an edit outside the expected surface, a network call, a merge. Agents coordinate with each other through bounded, authority-free messages and advisory file claims; none of that coordination can approve, dispatch, or widen anything on its own.

The other half is memory: a graph of decisions, constraints, conventions and attempts, fused at edit time with a code-level blast-radius analysis. An agent about to touch a symbol sees the prior, human-confirmed decisions anchored to it — and any contested proposals a human still needs to resolve — before it writes a line.

• Not a cloud service. There is no MUON account, no hosted brain, and no data leaves your machine except each vendor's own model traffic — which was already leaving; MUON does not add a hop.
• Not a replacement for your coding agents. It never ships or custodies one — you install and log in to each vendor CLI yourself, natively.
• Not fully hands-off by default. Approval gates are on unless you explicitly arm full-auto, and full-auto is a standing consent you grant per vendor lane, not a permanent setting.

Install the macOS app, then run a real mission end to end with the quickstart. Prefer a terminal? Everything the app does is reachable from the CLI, or from a session you launch yourself with MUON's MCP server attached.