Omnara

What Omnara does

Omnara runs AI agents on selected models and machines, with tools, secrets, permissions, and event streaming available through a managed API. It supports hosted use through Omnara Cloud and self-hosting under the Apache-2.0 license. Agents can execute on connected laptops, virtual machines, containers, or managed sandboxes, depending on the configured machine source.

The platform is aimed at applications and teams that need more than a one-off model request. A product can launch agents and send messages through the REST API or TypeScript SDK. Developers can also operate agents from the CLI, while teams can inspect tool calls and respond through the dashboard.

How it works

An agent configuration selects an instruction, model, and machine source. Omnara workers build model context, call the selected model, and execute the configured tools on the assigned machine. Supported model providers include OpenAI, Anthropic, OpenRouter, Amazon Bedrock, LiteLLM, Ollama, and compatible endpoints, including self-hosted models.

The API handles agent launches, user input, approval responses, and event streaming. PostgreSQL stores each agent step as its event log and state, allowing another worker to continue after a crash or restart. Workers are stateless and only handle an agent while it is taking a turn, so additional workers can process more turns concurrently. Machines connect outward through the omnarad daemon and do not require inbound ports.

Setup and configuration

Omnara Cloud provides an initial model provider and managed machine pool for a quick start. Launch an agent from an agent.yaml file with the Node.js CLI, which requires Node.js 22 or newer:

npx omnara agents launch --file agent.yaml \
  --name "Machine assistant" \
  --message "Report this machine's operating system and free disk space."

The TypeScript SDK requires an API key stored in OMNARA_TOKEN; the SDK can create an agent configuration, launch an agent, and consume its event stream. The CLI can also open an interactive session with a launched agent.

Self-hosting requires Docker with Compose. Clone the repository and start the application profile, then open the local console, add a model provider and model, connect a machine, and share both with a project. Configure the CLI with the local API and issuer URLs before signing in. Production deployments should use the self-hosting deployment and configuration guidance because local defaults are intentionally insecure.

Tools and capabilities

Agent configurations can include built-in tools for shell commands, file operations, web search and fetching, and asking a person a question. Omnara also supports attaching MCP servers and custom tools that run in external systems. Skills package instructions and files for reuse across agents.

Memory stores provide files that persist between conversations and can be shared with read or read-write access. Agents may create subagents with fresh context, inspect their progress, and provide additional instructions. Model credentials, tools, permissions, and machine assignments are configured as part of the platform rather than being tied to a single model vendor.

The deployment uses PostgreSQL for agent state and event history, Redis, blob storage for attachments and tool outputs, and a memory filesystem. The REST API, CLI, dashboard, and SDK use the same agent control surface, while events are available as the agent operates.