AI agents can already write code, search the web, and run commands. Keeping one running reliably for days is a different problem. A worker crashes, a laptop closes, a task waits for a human approval, and the agent loses its place. Omnara is an open source platform that turns agents into durable services. You launch an agent through an API. The platform keeps its state, runs it on any model and any machine, and lets it pick up exactly where it left off (omnara-ai/omnara, GitHub, retrieved 2026-10-25). The project reached a wide audience in the week of October 21, 2026, when its launch post reached the front page of Hacker News, and independent coverage followed on October 22 (Emergent). If you have ever babysat a long agent run, Omnara is built for that exact frustration.
Key Takeaways
Omnara is an open source managed agent platform under the Apache 2.0 license, launched publicly in the week of October 21, 2026 as a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API.
Agents are durable API objects: every step is saved in Postgres, so an agent survives crashes, restarts, and machine disconnects and resumes where it left off.
It is model agnostic and machine agnostic. You can use OpenAI, Anthropic, OpenRouter, or a self-hosted model, and run agents on your own laptop, VM, or one of eleven sandbox providers.
Control is built in: per-tool permissions that run immediately, ask a person first, or stay blocked, plus role-based access control, encrypted secrets, and human approval for sensitive actions.
The platform itself is free. You bring your own model keys and machines at no charge, or pay as you go with Omnara credits, or self-host the whole platform with Docker Compose.
The gap Omnara targets is familiar to anyone who has moved an agent from a demo to production. A demo agent answers one question and exits. A production agent works a queue of tasks for hours, waits for answers, and must not lose its place when something breaks. Hosted services solve this by running your agents on their infrastructure, but that ties you to one vendor's models, machines, cloud, and price list. Omnara's answer is to do the same job as open software: the control plane is yours to run, the models and machines are yours to choose, and the agent's data stays in your own database.
What Is Omnara?
Omnara is an open source managed agent platform: an API for running AI agents in production, with execution, state, tools, secrets, permissions, and streaming built in. That is how its GitHub repository describes it, and the word managed does most of the work. You do not write an agent as code and ship it inside your app. You launch an agent as an object through the API, then change it, message it, and watch it through the same API, without redeploying anything. The platform handles the parts every production agent needs and every team otherwise rebuilds: keeping state, running tools, asking humans for approval, and surviving interruptions.
The project is young as a managed agent platform, but it is not small. At the time of writing, the repository carries 630 commits, 3,233 stars, and 248 forks, and it is published under the Apache 2.0 license (omnara-ai/omnara, GitHub, retrieved 2026-10-25). Omnara began as a remote interface for coding agents and pivoted to agent infrastructure during 2026, so the managed agent platform itself is a recent product. That youth is worth keeping in mind alongside the enthusiasm of its launch week.
Aspect | Omnara |
|---|---|
Launched publicly | Week of October 21, 2026 (Show HN, front page) |
License | Apache 2.0 |
Repository | github.com/omnara-ai/omnara |
Positioning | Self-hostable alternative to Claude Managed Agents and OpenAI Agents API |
State storage | Postgres event log for every agent step |
Platform fee | $0. Bring your own models and machines, or pay as you go with credits |
How Omnara Works
An Omnara agent starts as a short YAML file. You write an instruction, pick a model provider and model name, and name the machines the agent may use. Then you launch it from the command line:
Define. Write agent.yaml with the instruction, the model, and the machine pool. A first project on Omnara Cloud already includes a model provider and a pool of managed sandboxes, so the example config runs as is.
Launch. Run npx omnara agents launch with the file, a name, and a first message. The CLI needs Node.js 22 or newer, and the same flow works from the TypeScript SDK or the REST API.
Talk to it. Chat with the agent in your terminal, or open it in the dashboard to watch its tool calls and reply. You can send new messages while the agent works, and events stream back as they happen.
Underneath, three parts do the work. The API accepts launches, messages, and approvals. Postgres holds every step of every agent as an event log, and that record is the agent's state, so any worker can pick an agent back up after a crash or restart. Stateless workers build the model context, call the model, and run tools, and an agent only occupies a worker while it is taking a turn, so one worker can serve many agents. The platform also uses Redis, blob storage for attachments and tool outputs, and a memory filesystem, according to the repository's architecture notes.
Machines are where agents touch the real world. A machine is a tool the agent uses. Install the Omnara daemon on any macOS or Linux laptop, VM, or container, and the agent can run commands and edit files there. No inbound ports and no SSH keys are needed, because the daemon connects out to Omnara. Or use sandboxes from eleven providers, including Daytona, E2B, Modal, and Blaxel. For example, one agent can use a laptop and a set of GPU servers at the same time, and move between them as the work changes. If your laptop goes offline, the agent can continue in a sandbox and sync its work back with git. That handoff, from a machine you own to a machine in the cloud and back, is the detail that makes durable agents practical rather than merely persistent.
Durable by Design: State, Schedules, and Subagents
Durable state is the event log itself: every step saved in Postgres, so any worker can rebuild the agent's position. Durability in Omnara is not a backup feature. It is the data model. Because every step lands in Postgres, an agent that crashes mid-task does not restart from scratch. It resumes from its event log, with its conversation, tool outputs, and approvals intact. Self-hosted operators get a useful side effect: every agent's history sits in their own database, where it can be queried for analytics, evaluations, and training datasets.
Three features extend that durability into everyday workflows. Schedules let cron triggers with time zones message an existing agent, launch a new one from a profile, or start a new Slack or Discord thread. Subagents let an agent start a fresh-context helper, check its progress, and send it new instructions, with each subagent fully inspectable. Memory stores give agents files that persist between conversations and can be shared with other agents, with read or read-write access per agent. Together they cover the three ways real agent work stretches over time: it repeats, it delegates, and it remembers.
Security and Control
Giving an agent a shell is a security decision, and Omnara treats it that way. Each tool can run immediately, ask a person first, or stay blocked, and the platform enforces the setting. When a tool needs approval, the agent waits until someone approves or denies the call. Role-based access control is how Omnara governs who can do what. Organization and project roles govern users and API keys. Credentials for models, machines, and MCP servers are stored once as encrypted secrets and shared with each project explicitly. One deployment can serve many organizations, each with its own members, projects, secrets, model providers, and machines.
The approval flow is also the human interface. On Slack or Discord, mentioning the bot starts an agent in a thread, where it replies, asks questions, and requests approvals. For example, on GitHub an agent can start when a pull request opens, read the diff, and submit a review. Agents that need a person do not fail silently. They wait, visibly, in a thread a human already watches.
Self-Host or Use the Cloud
Omnara gives you both doors. Self-hosting needs Docker with Compose: clone the repository, start the app profile, open the local console, and add a model provider and a machine. Local development prints the verification link in the Docker logs, so no email provider is needed. The repository is explicit that the local defaults are intentionally insecure, and a real deployment should follow its self-hosting guide and configuration reference. That honesty matters: a control plane that runs agents on your machines deserves a careful setup, not a one-command install left on defaults.
Omnara Cloud is the other door. New projects arrive with a default provider config pointing to OpenRouter and a pool of managed sandboxes, so the quickstart runs with nothing to configure. The platform itself charges no fee: bring your own model keys and machines at no charge, or pay as you go with Omnara credits for models at provider token rates and machines by active time. Idle agents cost nothing in either mode, because an agent only uses a worker while it is taking a turn.
What Independent Coverage Says
Launch-week coverage agrees on the shape of the product and on its main open question. Emergent stresses the cost angle: agents stored in Postgres survive restarts and cost nothing while idle, a sharp contrast with always-on implementations. StackFutures stresses the positioning: a self-hostable replacement for both Claude Managed Agents and OpenAI's Agents API, published under Apache 2.0 and launched to the Hacker News front page. Neither outlet claims production adoption figures for the new platform, and neither should: the managed agent platform is months old, and its public proof so far is the code, the documentation, and a launch that developers noticed.
The methodology of this article is source-first: every number here comes from the repository or a launch report cited beside it.
Who Should Use Omnara, and Who Should Not
Omnara fits teams that already feel the specific pain it removes:
Running many agents as a service, for customers inside a product or for internal teams, without redeploying on every change.
Keeping agents alive across crashes, restarts, and laptop disconnects, with a full event log to audit afterwards.
Avoiding lock-in to one vendor's models, machines, or cloud, while keeping agent data in a database they control.
Putting a human approval step in front of risky tool calls, in Slack, Discord, or GitHub threads.
Skip it if you want a single agent embedded in one app: an agent framework fits that job better, and Omnara's own site says so. Skip it, too, if you do not want to operate infrastructure and do not want a managed cloud either: the platform is free, but the models, machines, and the setup time are not. And if you self-host, budget the work of running Postgres, Redis, and blob storage properly. Durable agents are only as durable as the database underneath them.
Frequently Asked Questions
Is Omnara really open source? Yes. The platform is published at github.com/omnara-ai/omnara under the Apache 2.0 license, and it can be self-hosted for free. Omnara Cloud adds managed services on top of the same open code.
What happens when an Omnara agent crashes? It resumes. Every step of every agent is saved as an event log in Postgres, so after a crash, restart, or machine disconnect, a worker picks the agent back up from its saved state instead of starting over.
Which models and machines can Omnara use? Any model through OpenAI, Anthropic, OpenRouter, Amazon Bedrock, LiteLLM, Ollama, or a compatible endpoint you host yourself. Machines can be your own laptop, VM, or container through the Omnara daemon, or sandboxes from eleven providers including Daytona, E2B, Modal, and Blaxel.
How much does Omnara cost? The platform itself is free, with a $0 platform fee. You pay only for what you use: your own model keys and machines cost nothing extra, or pay as you go with Omnara credits for models at provider token rates and machines by active time. Idle agents cost nothing because they only use a worker while taking a turn.
How is Omnara different from an agent framework? A framework is a library for writing one agent as code and shipping it inside your app. Omnara is a control plane: agents are durable API objects that you launch, message, and change without redeploying, and the platform handles state, machines, permissions, and history for many agents at once.
The Bottom Line
Omnara turns the hardest part of production agents into infrastructure you can run yourself: state that survives crashes, machines that hand off between your laptop and the cloud, and approvals that keep a human in the loop. It is a young platform, and its launch-week attention is not yet proof of production adoption. But the design answers the right question. Agents are becoming services, and services need a control plane. Start with the GitHub repository and the official documentation. If you want to compare it with the evaluation side of agent tooling, our Microsoft Scope article covers the open platform that grades agents, and our OpenGAP article covers turning a git repo into a portable agent. Or read the Indonesian version of this article.