The open source framework for building agents.
Build contained, event-driven AI agents. Define your agent in a single config file, run it in a container and deploy it anywhere.
Open-source · Docker-native
waiting on triggers…
How it works
An agent is a file.
01Start with a purpose
Write a description in plain language of what your agents main job is.
Triggers & capabilities
Wire it into the tools your team already lives in.
Agents wake on the events you already produce — a webhook, a schedule, a message — and act through skills, tools, and tightly-scoped permissions.
Discord
type: discord
Slack
type: slack
Telegram
type: telegram
Webhook
type: http
Cron
type: cron
Skills
skills:
Tools / MCP
tools:
Memory
memory:
agent.yaml
One file defines your agent
Need a different trigger or integration?
The framework is open-source — open a PR on GitHub and wire in your own.
Security
Locked down, in nested layers.
An unattended agent has no one to ask “may I run this?” — so the answer is written down before it starts. Allowlists in the file, checked at three boundaries: the permission engine, the Deno runtime, and the container as the outer wall.
- Allowlists for net, run, read, and write — deny by default
- Every allow and deny logged to an audit trail
- Permissions change like code — file edit, review, redeploy
agent
Just use our image
Or use it as a base and build your own
FROM ghcr.io/loopedautomation/agent:latestRUN apk add --no-cache github-cliCOPY agent.yaml /agent/agent.yaml
Deployment
Runs anywhere.
Because every agent is a single container, you manage a fleet with the tools you already know — Docker, Compose, and whatever orchestrator you like. VPS, homelab, or cloud, it runs the same.
- docker run for one, docker compose for a fleet
- Runs on any VPS, homelab box, or cloud platform
- Fleet management through the container ecosystem you know
- Reproducible — the agent is the file plus the image
Once deployed
Agents are working.
Once deployed, each agent waits on its triggers and acts on its own — filing issues, running standups, paging on-call. You watch the stream; it does the work.
Live · Agent activity
Every run is logged, permission-scoped, and reproducible from the agent file.