Omnara is an open-source managed agent platform (Apache 2.0 license) enabling self-hosted AI agents compatible with Claude Managed Agents and OpenAI’s Agents API. It supports any model via OpenAI/Anthropic-compatible endpoints (e.g., OpenAI, Anthropic, OpenRouter) and runs agents on sandboxes from 11 providers (Arker, E2B, etc.) or user-controlled machines via the omnarad daemon. The architecture includes a REST API, TypeScript SDK, and CLI for launching agents defined in YAML (e.g., agent.yaml with model providers and machine pools). Agents execute tool calls (shell, files, web, MCP), persist state in Postgres, and resume after crashes. Features include subagents, memory stores, scheduling (cron triggers), and integrations with Slack/Discord/GitHub. Self-hosting requires Docker Compose, while Omnara Cloud offers free tier with user-provided keys or pay-as-you-go for platform services. Unlike closed alternatives, Omnara ensures data residency control and open-source extensibility. Setup involves cloning the repo, running `docker compose up -d`, and configuring models/machines. The system uses Redis, blob storage, and stateless workers to manage agent loops, with role-based access and secrets management. Example configs include PR review and browser agents, deployable via a single prompt.
Read original
hackernews