Omnara
Omnara · Omnara: Open-Source, Self-Hostable Alternative to Claude Managed Agents for Durable Agents on Any Model
Omnara is an open-source (Apache 2.0) managed agent platform: you define an agent in a YAML config, launch it through a REST API, TypeScript SDK, CLI or MCP, and Omnara runs the agent loop on the model you choose, with tools running on cloud sandboxes or your own laptops and servers. It suits product and platform teams that want hosted, long-running agents like Claude Managed Agents or OpenAI's Agents API without being tied to one model vendor or cloud.
Best for
Product and platform teams that want hosted, long-running agents behind an API, in Slack, or on GitHub, with the freedom to pick any model, run tools on their own machines or third-party sandboxes, and self-host under Apache 2.0 or keep agent history in their own Postgres
Not ideal for
Teams that want to run Claude Code or Codex themselves in the cloud, Python shops that need a native SDK, small teams unwilling to operate Postgres, Redis, and storage if they self-host, buyers who need a long vendor track record, and anyone looking for a ready-made agent rather than a platform
Who it's for
Product engineers embedding agents in their own app, platform and engineering teams running shared agents in Slack, Discord, or GitHub, and companies that want agent history and execution on infrastructure they control
Capabilities
- Agents are defined in a YAML config (instructions, model, tools, MCP servers, skills, memory stores, subagents, machine sources) and launched, messaged, and reconfigured through one REST API
- Durable state: every input, model output, and tool call is stored in a Postgres event log, so agents resume after crashes, restarts, and machine disconnects
- Stateless workers run one agent turn at a time and release the agent, so idle agents waiting on a person or a tool do not hold a worker; add workers to run more turns at once
- The `omnarad` daemon turns any macOS or Linux laptop, VM, or container into a machine an agent can run commands and edit files on, connecting outbound with no inbound ports or SSH keys
- Machine pools provision sandboxes on demand from providers including Arker, Blaxel, boxd, CreateOS, Daytona, E2B, Freestyle, Modal, Tenki, and Unikraft Cloud, and delete them when released or idle
- One agent can attach several machines and move between them, for example continuing in a cloud sandbox when a laptop goes offline
- Bring-your-own models through any OpenAI Responses, OpenAI Chat Completions, or Anthropic Messages compatible endpoint, with built-in presets for OpenAI, Anthropic, and OpenRouter and documented support for Amazon Bedrock, LiteLLM, Ollama, and self-hosted models
- Built-in tools for shell commands, files, web search and fetch, and asking a person a question, plus MCP servers and custom tools that execute in your own systems
- Per-tool permission modes (always allow, always ask, always deny) enforced by the platform, with the agent pausing until someone approves or denies
- Subagents with a fresh context that the parent can start, message, monitor, and stop, each inspectable as a full agent
- Memory stores: shared file stores attached to agents with read or read-write access that persist between conversations
- Slack and Discord bots that start an agent in a thread, and GitHub triggers that launch an agent when a pull request opens or someone mentions it
- Event streaming over SSE, signed event webhooks, and cron triggers with time zones that message an agent, launch a new one, or start a Slack or Discord thread
- Organization and project roles for users and API keys, shared secrets, multi-tenant organizations, and usage and cost reporting
- TypeScript SDK (`@omnara/sdk`), the `omnara` CLI, and a hosted MCP server that lets Claude Code, Codex, or Cursor launch and manage Omnara agents
Limitations
- Young product: Omnara relaunched as a managed agent platform in September 2026, and its server and daemon are still on frequent 0.x releases
- The only documented SDK is TypeScript; other languages use the REST API directly, and the CLI needs Node.js 22 or newer
- Production self-hosting means running the API, workers, maintenance service, and migrations against your own Postgres, Redis, S3-compatible storage, and a shared memory filesystem behind a public URL; the local Docker Compose defaults are documented as insecure
- A new self-hosted instance has no models or machines until you add a provider and connect a machine, and sandboxes require the instance to have a public URL
- Machines you connect yourself must run macOS or Linux; no Windows daemon is documented
- Omnara runs its own agent loop. No adapter for running existing harnesses such as Claude Code or Codex as the agent is documented
- Omnara-managed models and machines are usage-billed (token rates plus per-GiB machine runtime and retention), which is harder to forecast for long-running agents
- Compliance certifications are not listed in Omnara's docs or pricing page; security and compliance details sit in a separate Trust Center, so regulated buyers need to check there or ask the team
Use cases
- Embedding an agent in a SaaS dashboard that answers customer questions using your own tools, with each conversation's history kept in Postgres
- Running an internal coding agent in Slack that investigates an issue, works in a sandbox, and opens a pull request
- Reviewing every new GitHub pull request with an agent that checks out the repo, reads the diff, and leaves inline comments
- Giving an on-call SRE agent read-only access to AWS and your code to investigate alarms
- Scheduling a daily digest agent that summarizes posts about competitors or product usage and reports to a Slack thread
- Running an agent on your laptop while you work and letting it continue on a cloud sandbox when you close the lid
Our take
Omnara is one of a small group of open-source projects aiming to give teams the hosted-agent experience of Claude Managed Agents without the vendor lock-in, and its design choices are well documented: a Postgres event log as the agent's state, stateless workers that make idle agents nearly free, and an outbound daemon that lets the same agent use a laptop, a VM, or a third-party sandbox. Built-in Slack, Discord, and GitHub surfaces, approvals, subagents, and memory cover much of what in-house platforms otherwise build first. The pricing model is easy to like, since there is no platform fee and your own keys and machines are free. The caution is maturity. The platform relaunched in September 2026 and ships 0.x releases every few days, the SDK is TypeScript-only, and self-hosting in production is real infrastructure work. Teams evaluating it should compare it with Opengeni and TrueForge, which make similar promises, and confirm the security posture they need before putting customer data through it.
Who should use it
Teams building agents into a product or running shared internal agents who want an API-first runtime with durable state, approvals, and integrations; companies that need agents to work on their own machines or private networks; and anyone who wants Claude Managed Agents-style hosting with free model choice and the option to self-host.
Who should skip it
Developers who only need a coding agent on their own machine, teams that want Claude Code or Codex hosted as-is, organizations without the capacity to operate Postgres, Redis, and storage that still require self-hosting, and buyers who need a long-established vendor or must confirm compliance certifications before a trial.
Strengths
- Apache 2.0 with the same API, CLI, SDK, and config format on Omnara Cloud and self-hosted deployments
- Works with any model behind an OpenAI- or Anthropic-compatible endpoint, including self-hosted models
- Postgres event log and stateless workers make agents durable and cheap to keep idle
- Agents can use your own laptops, VMs, and containers without inbound ports, or sandboxes from many providers
- No platform fee on Omnara Cloud, with published per-GiB machine rates and free use of your own keys and machines
- Slack, Discord, GitHub, webhooks, schedules, approvals, subagents, and memory are built in rather than left to your code
Weaknesses
- Relaunched in September 2026 and still on 0.x releases
- TypeScript is the only SDK
- Self-hosting in production means operating Postgres, Redis, blob storage, and a shared filesystem
- Runs its own agent loop, so it does not host existing coding-agent CLIs
- Usage-billed managed machines and models can be hard to budget for long-running agents
Omnara pricing
Self-hosted
Free
- Full platform under Apache 2.0
- Run with Docker Compose or on your own infrastructure
- You pay only for your own infrastructure and model providers
Omnara Cloud (Self-serve)
$0 platform fee
- Full platform with no per-seat fees
- Bring your own model keys and machines for free
- Omnara-managed models at the provider's token rates
- Omnara-managed machines at $0.0414 per GiB of memory per hour online
- Machine retention at $0.20016 per GiB of memory per 30 days
- No credit card required to start
Enterprise
Custom
- Bring your own cloud or on-prem deployment with onboarding
- SSO, access controls, and audit-log setup
- Model fine-tuning on your data
- Custom integrations and branding
- Security reviews, custom SLAs, and priority founder support
Free tier limits: Self-hosting is free. On Omnara Cloud the platform is free; only Omnara-managed models and machines are billed
Note: Idle Omnara-managed machines auto-sleep, which pauses runtime charges while keeping their data; retention is billed in addition to runtime. Omnara's example: a 4 GiB machine active one hour a day costs about $5 a month in runtime plus $0.80 in retention. Accepted startups can get credits for models and machines through Omnara for Startups.
Technical specs
API pricing
API access is free on Omnara Cloud and self-hosted; Omnara-managed models are billed at provider token rates and Omnara-managed machines per GiB of memory per hour online plus retention
Available models
Where Omnara excels
Agents inside a SaaS product
The REST API, TypeScript SDK, event streaming, and per-tool approvals let a product team launch an agent per customer request and show its progress in their own UI, with history stored in Postgres.
Team coding agent in Slack
Mentioning the bot starts an agent in a thread that works in a sandbox, asks for approvals there, and can open a pull request, without each engineer running an agent locally.
Agents that need private infrastructure
Installing omnarad on a VM inside a private network or a GPU server gives agents access to internal systems and hardware without opening inbound ports.
Leaving a hosted agent platform tied to one model vendor
Teams that want the Claude Managed Agents pattern but with OpenAI, open-weight, or self-hosted models, or with data kept in their own database, can self-host Omnara or use its cloud with their own keys.
Omnara vs. competitors
Omnara vs. Opengeni
Opengeni is also an Apache-2.0 runtime for long-running agents with a Postgres event log, approvals, and enrolled machines, but self-hosting it involves Temporal and NATS and its hosted cloud charges model cost plus 5%. Omnara's stack is Postgres, Redis, storage, and Go workers, its cloud has no platform fee, and it ships built-in Slack and Discord bots alongside its GitHub triggers.
Omnara vs. TrueForge
TrueForge is TrueFoundry's MIT-licensed agent harness, also pitched as an alternative to Claude Managed Agents, with a bundled chat UI and Helm chart, and it currently relies on Daytona for sandboxes. Omnara offers many sandbox providers plus your own machines through its daemon, and a hosted cloud with published usage rates.
Omnara vs. DigitalOcean Managed Agents
DigitalOcean Managed Agents hosts existing harnesses such as Claude Code, Codex, and OpenCode in Firecracker microVMs on DigitalOcean's cloud, with a 16,000-tool MCP gateway, and cannot be self-hosted. Omnara runs its own model-agnostic agent loop, can be self-hosted, and lets agents use machines anywhere.
Omnara vs. ZooWork
ZooWork pairs a no-code Agent Builder with a Managed Agent API in developer preview that runs agents in E2B sandboxes on its own cloud. Omnara is developer-first, open source, and self-hostable, with a choice of models and machines.
Frequently asked questions
What is Omnara?
Omnara is an open-source managed agent platform. You define an agent in a config file, launch it through an API, and Omnara runs the agent loop on the model you choose, saves every step to Postgres, and gives the agent tools, sandboxes or your own machines, approvals, memory, and integrations with Slack, Discord, and GitHub. It positions itself as a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API.
Is Omnara open source?
Yes. The full platform is in the omnara-ai/omnara GitHub repository under the Apache License 2.0, and published container images for the web app, API, worker, maintenance service, and migrations accompany each release.
How much does Omnara cost?
Self-hosting is free. Omnara Cloud has no platform fee or per-seat charge: you can bring your own model keys and connect your own machines for free, or pay as you go for Omnara-managed models at the provider's token rates and Omnara-managed machines at $0.0414 per GiB of memory per hour online plus $0.20016 per GiB per 30 days retained. Idle machines auto-sleep. Enterprise deployments with on-prem or bring-your-own-cloud setup, SSO, and custom SLAs are priced by contact, and a startup program offers credits.
Can Omnara be self-hosted?
Yes. The repository includes a Docker Compose file that runs the whole stack on one machine for local use. Production deployments run the API, workers, and maintenance services plus a one-shot migration job against your own Postgres, Redis, S3-compatible storage, and a shared memory filesystem, with a public URL in front of them.
Which models does Omnara support?
Any model behind an OpenAI Responses, OpenAI Chat Completions, or Anthropic Messages compatible endpoint. Built-in presets cover OpenAI, Anthropic, and OpenRouter; Amazon Bedrock (including AWS credentials), LiteLLM, Ollama, vLLM, and other self-hosted or proxy endpoints are configured with a custom base URL and API format. On Omnara Cloud, your first project comes with an Omnara-managed OpenRouter provider billed at token rates and a pool of managed sandboxes.
Where do Omnara agents run their commands?
On machines. You can install the omnarad daemon on any macOS or Linux laptop, VM, or container, which connects outbound so no inbound ports are needed, or let Omnara create sandboxes on demand from providers such as Daytona, E2B, Modal, Blaxel, or Unikraft Cloud. An agent can use several machines and switch between them.
Omnara vs Claude Managed Agents: what is the difference?
Both run hosted agents you create and message through an API. Claude Managed Agents is a closed service on Anthropic's cloud that runs Claude models with Anthropic's loop. Omnara is Apache 2.0, works with any compatible model, can be self-hosted or used on Omnara Cloud, and stores agent history in Postgres you can query when you self-host.
Is Omnara still a mobile app for Claude Code?
Omnara's early product was a remote-control phone app for Claude Code and Codex. The company has since built a different product, a cloud platform where agents run as a service and any client talks to them through an API, and that platform is what this page covers.
Integrations & fit
Alternatives to consider
Cloudgeni
Product and platform teams building agent features into their own TypeScript/React product, or running agents for their organization, who want durable sessions, approvals, credential handling, and the choice to self-host under Apache-2.0 or start on a hosted cloud with the same API
Freemium
TrueFoundry
Platform and application teams that want a model-neutral, self-hosted runtime for reusable production agents
Free
DigitalOcean
Developers and small teams who already use Claude Code, Codex, OpenCode, Hermes, or LangGraph and want those agents running durably in the cloud, with pause, resume, and fork, governed tool access, and usage-based billing in one place
Paid
Serendipity One
Consultants, agencies, forward-deployed engineers, and operations teams who want to package a repeatable workflow as an agent with approvals and schedules, deliver it to several teams or clients, and optionally embed it in their own product through an API
PaidAbout Omnara
Omnara started as a remote-control phone app for Claude Code and Codex, then relaunched in September 2026 as an open-source platform for running agents as a service. You write an agent config (instructions, model, tools, MCP servers, skills, memory stores, subagents, and which machines it may use), launch it with one API call, then message it, stream its events, and answer its approval requests from your app, the CLI, the dashboard, Slack, Discord, or GitHub. The architecture is what sets it apart from local harnesses such as Claude Code or OpenCode. Each agent's history lives as an event log in Postgres, and stateless Go workers run one turn at a time and release the agent, so agents survive crashes and restarts, hold no worker while they wait on a person or a tool, and scale by adding workers. Tool execution happens elsewhere: any macOS or Linux machine running the small `omnarad` daemon connects outbound to Omnara with no inbound ports, or Omnara provisions sandboxes on demand from providers such as Daytona, E2B, Modal, Blaxel, and Unikraft Cloud. One agent can use several machines and keep working in a sandbox when your laptop goes offline. Models come from your own keys for OpenAI, Anthropic, OpenRouter, Amazon Bedrock, LiteLLM, Ollama, or any compatible endpoint. Self-hosting is free under Apache 2.0, starting with Docker Compose on one machine. Omnara Cloud has no platform fee: bring your own keys and machines for free, or pay provider token rates for Omnara's models and $0.0414 per GiB-hour for its managed machines. The tradeoffs are youth and scope. The relaunched product is weeks old and still on 0.x releases, the only SDK is TypeScript, and production self-hosting means operating Postgres, Redis, blob storage, a shared memory filesystem, and a public URL. Omnara also runs its own agent loop, so it is not a host for existing coding-agent CLIs.
Updates from Omnara
Omnara posted a Show HN on October 21, 2026 presenting itself as an open-source alternative to Claude Managed Agents, describing its Postgres event log and stateless-worker design. The same day's 0.2.9 release of the server and the omnarad daemon added E2B as a machine pool provider.
Omnara added Slack and Discord bots that start agents in threads, GitHub triggers on pull requests and mentions, shared memory stores agents can read and write, a redesigned dashboard, deployable example agents, and Tenki, Arker, boxd, and CreateOS as sandbox providers.
Agents can now spawn subagents with a fresh context, send them messages, read their progress, and stop them. The release also added deferred tool loading with tool search, usage and cost pages, and custom model endpoints in the dashboard.
A hosted MCP server at app.omnara.com/mcp lets Claude Code, Cursor, or other MCP clients launch and manage Omnara agents, alongside the API, SDK, and CLI. Modal sandboxes and Amazon Bedrock support arrived in the same update.
Omnara relaunched on September 2, 2026 as an open-source alternative to Claude Managed Agents, with durable agents, a choice of models, Omnara-managed or bring-your-own machines, MCP, custom tools, and skills.
Are you the founder? Claim this listing →