Skip to main content
Omnara is the open-source managed agent platform: the self-hostable alternative to Claude Managed Agents and OpenAI’s Agents API. Launch production AI agents through one API, on any model and any machine, with tools, secrets, permissions, approvals, and streaming built in. Use Omnara Cloud, where there’s nothing for you to operate, or self-host the whole platform for free under Apache 2.0. Unlike the AI labs’ managed agents, Omnara runs any model, and you can run all of it yourself with agent history in your own Postgres. Unlike frameworks such as LangGraph or Mastra, it’s a running platform, not a library you host and build around.

​
Use cases

Omnara can be used to build agents for internal use cases or external use cases, such as:
  • Internal Slack bots
  • Code review agents
  • Agents integrated in a product dashboard
Omnara exposes a web interface, CLI, and REST API to enable these use cases.

​
How it works

You describe an agent in a config: its instructions, model, tools, and machines. Omnara runs the agent loop on its own workers and saves every step to Postgres as it happens.
  • Agents are durable. After a worker crash, a restart, or a machine disconnect, an agent picks up from its last saved step, and recovery doesn’t rerun tools. Failure handling covers each case.
  • Scaling is adding workers. Workers are stateless, and an agent only uses one while it’s taking a turn.
  • Agents are API objects. Launch one per customer, ticket, or thread. Post a new config to change a running agent’s model, tools, permissions, or instructions at its next model call, or send a steering message, without restarting the conversation or redeploying.
  • Your logic stays in your code. Custom tools run in your own backend, in any language, and Omnara holds each call until you answer. MCP servers work too.
  • Per-tool permissions. Each tool can run immediately, require approval, or be blocked, and the platform enforces it.
  • Configs are files. Keep a config in your repository, review it in pull requests, and apply it from CI. Every change creates a new version.

​
Build an agent

An agent starts as a small YAML config:
You can create and launch the agent in two ways:
  • Dashboard: click New agent in your project in the dashboard.
  • Programmatically: Omnara provides an API, TypeScript SDK, and CLI to interact with Omnara.
The quickstart provides a more detailed guide to setting up your first agent.

​
What’s included

  • Any model: OpenAI, Anthropic, OpenRouter, Amazon Bedrock, or any compatible endpoint, including models you host, on your own keys.
  • Machines: your own computers, VMs, and containers through a daemon that only makes outbound connections, or sandboxes from providers like E2B, Daytona, and Modal. An agent can use several at once.
  • Built-in tools: shell commands, files, web search and fetch, and questions for a person.
  • Secrets: encrypted and shared with projects through grants. MCP credentials stay on Omnara’s workers, never on the agent’s machines.
  • Skills, memory, and subagents: subagents you can inspect, message, and stop, with limits on how many and how deep.
  • Schedules, integrations for Slack, Discord, and GitHub, streaming that resumes after a reconnect, and webhooks.

​
Open source and Omnara Cloud

The whole control plane, including the agent loop, recovery, and the tests that cover them, is in the omnara-ai/omnara repository under Apache 2.0, and the self-hosting guide covers a production deployment. Omnara Cloud runs the same control plane with no platform fee: you pay provider token rates for models bought through Omnara and usage rates for Omnara-managed machines, and nothing for your own keys and machines. Omnara’s team can also deploy and manage it in your own cloud (BYOC) or on premises, with SSO, audit logs, and custom SLAs: book a call with the founders. See pricing and the Trust Center.

​
Core concepts

  1. An organization owns shared resources for your team: projects, secrets, model providers, and machines.
  2. Projects contain agents. Grants decide which shared resources the project can use.
  3. An agent is one durable conversation with its config, tool calls, and history. Profiles provide reusable agent configurations.
  4. You send inputs, stream events, and resolve interactions in an agent.
The concepts page gives a short overview of this model.

​
Start here

Quickstart

Write a config, create the grants, and launch an agent with the dashboard, the API, or the CLI

Concepts

The object model, the agent loop, and access control

Failure handling

What happens to a running agent when a worker, model call, tool, or machine fails

Self-hosting

Run the whole platform on your own infrastructure