Omnara

Overview

Omnara is an open-source managed agent platform for running AI agents in production. It provides an API that handles execution and state while letting teams choose models, tools, machines, and how users interact with each agent. It is available through Omnara Cloud or as a self-hosted deployment.

Key Features

• Durable agents store state in a durable event log, can run indefinitely, and use no compute while idle.
• Run agents on managed sandboxes, bring your own Blaxel, Daytona, or Unikraft Cloud account, or install the Omnara daemon on a laptop, VM, or container; the daemon connects outbound.
• Access controls include human approval for risky tool calls, encrypted scoped secrets, role-based access for people and API keys, and an append-only audit history.
• Agent-ready APIs include a REST API, TypeScript SDK, and CLI for launching, steering, and observing agents and for changing their config mid-conversation without redeploying.
• Model-agnostic support includes OpenAI, Anthropic, OpenRouter, LiteLLM, Amazon Bedrock, Ollama, and compatible endpoints, including self-hosted models.
• Integrations connect MCP servers and reusable skills and let teams talk to agents through first-party Slack, Discord, and GitHub integrations.

Use Cases

• Agents embedded in a product dashboard.
• Internal Slack agents for a team.
• Specialized agent workflows such as code review, SRE investigation, browser operations, and video generation.
• Interactive CLI experiences.

Getting Started

• Sign in to Omnara Cloud; the first project includes a model provider and a default sandbox pool.
• Save an agent configuration as agent.yaml and launch it with the Omnara CLI; Node.js 22 or newer is required.
• Chat with the agent in the terminal using npx omnara agents chat <agent-id>, or open it in the dashboard to watch tool calls and reply.
• For programmatic use, create an API key, export it as OMNARA_TOKEN, install @omnara/sdk, and use the org and project IDs printed by npx omnara config.

Deployment & Requirements

• Self-hosting requires Docker with Compose.
• Start the app profile with docker compose -f compose.yaml --profile app up -d from a cloned repository.
• A new self-hosted instance has no models or machines; add a model provider and model, connect a computer under Machines, and share both with the project.
• Sandboxes require the self-hosted instance to have a public URL.
• Local defaults are intentionally insecure; real deployments should follow the self-hosting and configuration documentation.

Before You Adopt

• License: Apache-2.0. Review its terms before using, modifying, or distributing the project.
• Self-hosted local defaults are intentionally insecure and should not be used as-is for a real deployment.
• A new self-hosted instance starts without any models or machines, so these must be added before agents can run.
• Managed agent platforms can lock models, machines, cloud, data, and pricing to one vendor.