Archon

The control room for your AI agent fleet.

Most agent tooling only watches. Archon also gives you the controls — kill switches, start/stop, restart — so when an agent or automation goes wrong, you can actually stop it. One live pane for everything your agents are doing, plus a control API to act on it.

Why Archon

Running several agents (OpenClaw, Claude Code, CrewAI) plus automations (Make/n8n) means events scattered across a dozen logs. Archon normalizes them into a single event stream, streams it live to one dashboard, and — critically — lets you act: hit a kill switch, restart a service, stop an automation.

• See — every agent session, automation, and service in one live feed
• Control — kill switches, start/stop/restart, pre-configured and safe
• Alert — Telegram/webhook on failures and state changes
• Zero dependencies — a stdlib Python server + one HTML file

Quick start

git clone https://github.com/archonhq/archon && cd archon
pip install .
archon serve --sessions-dir ~/.openclaw-autoclaw/agents/main/sessions \
             --control-config examples/control-config.json
# open http://127.0.0.1:8000

How it works

framework logs ──► adapter ──► event stream ──► dashboard (HTTP/SSE)
                                    ▲
                               control API
                                    │
framework APIs ◄─────── executor ◄──┘

Adapters translate framework-specific logs into one JSON event stream (see the telemetry contract). The dashboard and control API consume only that contract — a new framework is one adapter away.

Control is safe by design. Commands are pre-configured in a config file, never supplied by the client. The control API accepts only {action, target} and routes it to a matching command — client input never reaches the shell unbounded. See examples/control-config.json.

Adapters: OpenClaw (reference) · Make.com · trading monitor · Claude Code (planned) · CrewAI (planned) · n8n (planned) · generic webhook (planned)

Status

Pre-alpha. Reference implementation: OpenClaw adapter. The control plane is proven in production on a live trading-monitor deployment (kill switch, risk rails, Telegram alerts, VPS health).

Roadmap

License

MIT