Omnara

Omnara is an open-source platform for running AI agents as a hosted service. You define an agent in a YAML config, launch it through the REST API, TypeScript SDK, CLI, or dashboard, and Omnara runs the agent loop: building the model context, calling the model, executing tools, and streaming every event back to your app. Every step is written to Postgres, so an agent survives worker crashes, restarts, and machine disconnects and picks up where it left off.

Agents can use any OpenAI Responses, OpenAI Chat Completions, or Anthropic Messages compatible model, including OpenAI, Anthropic, OpenRouter, Amazon Bedrock, LiteLLM, Ollama, and self-hosted endpoints. Tools are pluggable: built-in shell, file, web search/fetch, and ask-a-human tools, plus MCP servers and custom tools that call into your own systems over webhooks. Skills package reusable instructions and files, memory stores give agents files that persist between conversations, and subagents run in fresh contexts that you can inspect. Each tool can be set to run immediately, require human approval, or be blocked.

Compute is separate from the platform. Install the omnarad daemon on a macOS or Linux laptop, VM, or container and agents can run commands there with no inbound ports or SSH keys, or use sandboxes from providers such as Daytona, E2B, Modal, Blaxel, and Freestyle. A single agent can use several machines at once. Omnara integrates with Slack, Discord, and GitHub, supports cron triggers with time zones, and exposes streaming events and webhooks. Teams get organizations, projects, role-based access control, and shared secrets.

Self-hosting is a Docker Compose deployment (API, workers, Postgres, Redis, and blob storage) under the Apache 2.0 license. Omnara Cloud is the managed option, free when you bring your own model keys and machines and pay-as-you-go for Omnara-hosted models and sandboxes.