Agent Barn
Open-source control plane for AI agents on your own Kubernetes cluster. Agents work in Slack, Teams, Telegram and Discord, with per-agent costs, access controls and audit logs.
Overview
Description
Agent Barn is an open-source (Apache-2.0) control plane for AI coworkers that run on your own infrastructure. You define an agent once, connect it to the tools your team already uses, and it lives in Slack, Microsoft Teams, Telegram or Discord. Each agent runs in its own pod in your Kubernetes namespace, and every conversation, tool call and dollar of model spend is recorded. It runs agent runtimes such as Hermes and OpenClaw, with no seat licence.
Features
- Chat Integrations: Agents work in Slack, Microsoft Teams, Telegram and Discord, with per-agent channel allowlists.
- A Model Per Agent: Each agent gets its own virtual key on a LiteLLM proxy you operate.
- Cost Attribution: Per-agent and per-organisation token and spend tracking instead of one opaque bill.
- Access Control: Multi-tenant organisations with permission-backed roles and per-agent access roles.
- Audit Trail: Conversation history, a tool-call log and a domain-event security audit.
- Integrations & Skills: Google Workspace, Jira, GitHub, Confluence and custom APIs, plus versioned, forkable skills.
Technology Stack
The API and workers are written in Python with Dramatiq and PostgreSQL, the admin UI in TypeScript, and model traffic is routed through LiteLLM. Deployment uses Helm charts and Helmfile, with namespace-scoped Prometheus and Grafana.
Requirements
A Kubernetes cluster for production. Locally, Docker with Compose v2.24+ starts the full stack, including a k3d cluster. You also need an OpenRouter or other model provider key.