Omnara
Omnara is an open-source API and control plane for running production AI agents: durable execution that survives crashes, real machines as tools, human approvals, encrypted secrets and append-only audit. It is model-agnostic, self-hostable, and free to start.
What is Omnara?
Omnara is an open-source API and control plane for running AI agents in production. Its premise, stated in the documentation, is that an agent a product or a team can rely on is a service: it needs hosting, user and team access, integrations, durability, and clear rules about what it is allowed to do. Omnara supplies that layer, leaving you to build only the instruction, your own tools and integrations, and your product interface.
The object model is deliberately small. An organisation owns the shared resources — projects, secrets, model providers and machines. A project scopes agents and the grants deciding which shared resources they may use. An agent is one durable conversation with its config, its tool calls and its history. The workflow follows three steps: define an agent from the console or a short YAML config, launch it over a REST API, then converse with it — streaming its events, sending follow-up input, and resolving the approvals and questions it raises.
Durability is the core claim. Agents run for minutes, hours or days, survive crashes and resume exactly where they stopped, because every turn, tool call and approval is durable state rather than memory inside a process. Agents also get real computers as tools: managed pool machines they create and destroy on demand, or your own hardware connected through a single outbound daemon, with no SSH or port forwarding. Around that sit envelope-encrypted secrets, MCP integrations, versioned Skills, checkpoints, artifacts, organisation and project roles, and an append-only audit history enforced by the database. Tool policies — always_allow, always_ask, always_deny — are applied by the platform rather than by the prompt.
Omnara is model-agnostic: OpenAI, Anthropic, OpenRouter or any compatible endpoint, including models you host yourself. The platform is published under Apache-2.0 and the managed cloud runs the same code, so a self-hosted install in your own VPC exposes an identical API.
Alongside it, Omnara ships an earlier product: a command center for Claude Code and Codex with desktop, web, iOS, Android and Apple Watch clients, parallel agents, worktrees and Git integration. API access itself is still gated behind an early-access waitlist.
What it does
- Define an agent — instruction, model, tools, policies and machines — from the console or a short YAML config
- Launch it over a REST API and stream its events live on a resumable feed
- Send follow-up input mid-run, and interrupt or steer an agent that is already working
- Give agents real computers: managed pool machines on demand, or your own hardware through one outbound daemon
- Gate risky tool calls behind always_allow, always_ask or always_deny policies enforced by the platform
- Resolve approvals and questions from Slack or from the console
- Query an append-only audit history of every input, model call, tool call and approval
When to use Omnara / When not to
A quick filter to help you decide if Omnara is the right fit.
When to use Omnara
- Engineering teams shipping a customer-facing agent product who do not want to rebuild durability, streaming, approvals and audit themselves
- Platform and DevOps engineers who need agents to run on real machines with scoped credentials and enforced tool policies
- AI and LLMOps engineers standardising how agents are defined, launched, observed and replayed across projects
- Security-conscious organisations that require self-hosting in their own VPC or on-premises, with SSO and a SIEM-wired audit trail
- Individual developers who want to steer Claude Code or Codex from a phone, browser, desktop or Apple Watch
When not to use Omnara
- Non-technical users looking for a no-code agent builder: everything goes through a REST API, the console or a YAML config
- Teams wanting a ready-made vertical agent for support or sales: Omnara is infrastructure to build on, not a finished assistant
- Buyers who need a signed DPA, a subprocessor list or a SOC 2 report before a pilot, since none of them is published
- Organisations with strict data-residency obligations on managed cloud, because no hosting country or region is disclosed
- Anyone who needs access immediately: the API product is still gated behind an early-access waitlist
How to use Omnara
A typical end-to-end flow, from setup to results.
- Join the waitlist from the homepage, since API access is still rolling out as early access
- Read the quickstart in the documentation, which targets a first agent launched and streaming in about ten minutes
- Create an organisation and a project, which scope your secrets, model providers and machines
- Register a model provider — OpenAI, Anthropic, OpenRouter or any compatible endpoint — with your own key, or top up managed credits
- Define an agent in the console or in a short YAML config: instruction, model, allowed tools and policies
- Attach machines: create a managed pool, or connect your own machine with the outbound daemon, without SSH or port forwarding
- Authenticate against the API and launch the agent, which starts working immediately
- Stream its events, send follow-up input, and resolve approvals and questions in Slack or the console
- Extend it with MCP servers, custom tools and versioned Skills as the use case grows
- Or self-host the same platform in your own VPC, and install the desktop app or CLI to steer Claude Code and Codex from any device
Pros & Cons
Pros
- The source code is public under Apache-2.0, in an active repository of 2,714 stars and 207 forks pushed on the day of this review
- The managed cloud runs the same code as the self-hosted build and exposes an identical API, which limits lock-in
- No per-token markup and no per-seat fee: bring your own model keys, or top up credits at provider list rates, and credits do not expire
- Connecting your own machines is unlimited and free of charge
- Data stays queryable and exportable as Postgres rows, and never leaves your infrastructure when self-hosted
- Extensive public documentation: 27 guide pages plus an API reference covering more than 120 endpoints
- A companion command center for Claude Code and Codex already ships on the App Store and Google Play, announced as free with unlimited sessions
Cons
- API access is still behind a waitlist, so most of what the homepage describes cannot be tried freely today
- No compliance documentation at all: no GDPR mention, no Data Processing Agreement, no subprocessor list, no Article 27 representative
- No security certification is claimed anywhere — no SOC 2, no ISO 27001, no HIPAA
- No postal address and no legal form are published; the company simply calls itself Omnara
- The published privacy policy and terms were written for the mobile application, not for the API platform
- No hosting country or region is disclosed for the managed cloud
- A four-person company founded in 2025, and the Custom plan price is not public
Pricing & Plans
Omnara offers a permanent free plan at USD 0, including a basic VM, USD 10 of model credits, unlimited bring-your-own machines and no credit card requirement. Beyond that entry point, the lowest published price is USD 10 per machine per month, with model credits topped up at provider list rates and no per-seat fee. The Custom plan, intended for deployment inside the customer's own environment, is quoted on request. Self-hosting the open-source platform carries no licence cost.
- basic VM included
- USD 10 of model credits
- top-ups at provider list rates
- unlimited bring-your-own machines
- your own model keys
- no credit card required
- deployment in your VPC or on-premises
- SSO (OIDC including Microsoft Entra ID) and RBAC setup
- audit trail wired into your SIEM
- a queryable Postgres database you own
- self-hosted and open-source model setup
- versioned releases with change control
- security review support
- and a direct founder channel with committed response times
- the platform is open source under Apache-2.0 and runs in your own infrastructure
Data, GDPR & hosting
A consolidated view of how Omnara handles your data.
GDPR overview
There is no GDPR mention anywhere. Searching every page collected — homepage, pricing, documentation, privacy policy and terms — returns zero occurrences of GDPR, and none of CCPA either. No EU representative under Article 27 is designated, no data protection officer is named, no Data Processing Agreement is published or offered, and no subprocessor list exists. The rights the privacy policy does list — access, deletion, correction, portability, opting out of the sale or sharing of personal data, non-discrimination — follow Californian wording rather than European wording, and the terms designate the laws of the United States as governing law without naming a state. The only channel for exercising a right is the general contact address. Self-hosting on the Custom plan is the single documented answer to data-residency constraints.
Who owns the data?
Omnara's terms state that you retain ownership of the content you create through the service, and that Omnara takes only the licence needed to store and display it while providing the service. API keys and credentials you supply stay your responsibility. The vendor's own framing is that an agent is its event log: an append-only history of every input, model call, tool call and approval, stored as Postgres rows you can query and export, and which never leaves your infrastructure when you self-host. The privacy policy adds that personal information is neither sold nor shared, and is disclosed only to service providers operating the platform, in aggregated or de-identified form, or where the law requires it.
Reuse rights
Nothing in the terms restricts what you do with your own content or with an agent's output: you keep ownership, and no permission is required to reuse, export or redistribute your event log, which the vendor explicitly presents as queryable and exportable Postgres data. On Omnara's side, the privacy policy limits use to operating and improving the service, processing sandbox credits, relaying messages between you and your agents, sending notifications, answering support requests, analysing usage and meeting legal obligations. It states that personal information is neither sold nor shared. It says nothing at all about using customer data to train models, in either direction.
Data retention & training
Hosting summary
Omnara does not disclose where customer data is hosted on its managed cloud: no country and no region is named on the site, in the documentation, in the privacy policy or in the terms. The website itself resolves to a Cloudflare anycast address, which says nothing about where agent data lives. What the vendor does document is the alternative: the Custom plan deploys the whole platform — API, workers, console and database — inside your own VPC or on-premises, shipped as versioned container images that run anywhere, including your Kubernetes clusters, with a queryable Postgres database you own. In that configuration the event log never leaves your infrastructure. Secrets are described as envelope-encrypted with AES-256-GCM and scoped by organisation, project and user. Governing law is stated as that of the United States, without naming a state. Anyone with data-residency obligations should treat self-hosting as the only documented answer and ask directly about the managed cloud.
Things to keep in mind
Risks and trade-offs to weigh before adopting Omnara.
- Agents run for hours or days on real machines with real credentials: a careless instruction or an over-permissive policy can do damage fast, and only the approval rules you configure stand in the way
- Tool policies and approvals are only as good as their configuration; setting everything to always_allow removes the very safeguard the platform exists to provide
- Delegating long stretches of work to autonomous agents makes it easy to stop reading what they actually did, and an audit log only helps if someone reviews it
- The published privacy policy and terms cover the mobile application rather than the API platform, so the contractual perimeter of the product described here is unclear
- The liability cap in the terms is limited to what you paid over the past twelve months and explicitly excludes data breaches, unauthorised account access and data loss
- No GDPR framework, Data Processing Agreement, subprocessor list or security certification is published, which will stall procurement in regulated environments
- Model credits, machines and provider keys are all handled in one account, so a single compromise exposes credentials and running work at once
Setup & Integrations
Technical difficulty
Developer-level. Integration runs through a REST API and a resumable event stream, and agents are defined in the console or in a short YAML config, with no no-code path. The documentation targets a first agent launched and streaming in about ten minutes from the quickstart. Connecting your own machine takes a single outbound daemon, without SSH, port forwarding or tunnelling. Self-hosting means running versioned container images, Kubernetes included, and the Custom plan bundles on-site deployment help. The companion command center is far simpler: download the desktop app or install the CLI.
Deployment
Apps stores
Integrations
Behind Omnara
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Omnara.
Frequently asked questions
What exactly is Omnara?
What do I still have to build myself?
Do I need to run any infrastructure?
Which models can I use?
Is it genuinely open source?
How much does it cost?
How do humans stay in control of an agent?
Can I use it right now?
What does Omnara say about the GDPR?
Who is behind the product?
Should you pick Omnara?
Omnara is a real and unusually verifiable product for a company this young. The platform it describes is published under Apache-2.0 in a repository of more than 2,700 stars that was pushed on the day of this review; the documentation runs to twenty-seven guide pages plus an API reference of more than a hundred and twenty endpoints; the companion command center for Claude Code and Codex is live on the App Store and Google Play; and the pricing page carries actual figures instead of a contact form. Behind it is a four-person team from the Y Combinator Summer 2025 batch, with both founders publicly named.
The technical proposition is coherent. Durable execution, machines as tools, platform-enforced approval policies and an append-only audit log are precisely the pieces teams end up rebuilding themselves, and the homepage argues that case with public quotes from engineering teams at Sentry, Stripe, Shopify and others. Because the managed cloud runs the same open-source code and exposes the same API as a self-hosted install, the usual lock-in objection largely falls away.
The gap is compliance, and it is wide. There is no mention of the GDPR anywhere, no Data Processing Agreement, no subprocessor list, no EU representative, no security certification, no disclosed hosting location, and no postal address or legal form. The privacy policy and terms that do exist were written for the mobile application rather than the API platform, so the contractual perimeter of the product on offer is unclear. Access is also still gated behind a waitlist.
For an engineering team able to self-host and read the code, Omnara is worth a serious look today. For a regulated buyer who needs paperwork before a pilot, the paperwork is not there yet.
- Choosing a selection results in a full page refresh.
- Opens in a new window.