Open-source Omnara pitches self-hosted managed agents against OpenAI and Anthropic APIs

An open-source project named Omnara has landed on Hacker News's front page with a pitch aimed squarely at the managed agent services from Anthropic and OpenAI. According to the project's GitHub README, Omnara is a managed agent platform that developers can self-host for free under an Apache 2.0 licence, or run through the vendor's hosted Omnara Cloud, and it is positioned as an alternative to Claude Managed Agents and OpenAI's Agents API.

How the platform works

The README describes a system built around four pieces. A single REST API — also used by the CLI and a web dashboard — handles launching agents, sending messages and answering approval requests, with events streamed back as they happen. A Postgres database records every step each agent takes; that log is treated as the agent's state, so any worker process can pick the agent back up after a crash or restart. Workers themselves are stateless processes that assemble the model context, call the model and run tools, occupying a worker only for the duration of a turn, which lets one worker serve many agents. Finally, machines run a daemon called omnarad that dials out to the platform, so agents can execute commands on a laptop, VM or container without opening any inbound ports. The stack also depends on Redis, blob storage for attachments and tool outputs, and a memory filesystem.

Models, tools and agent features

Omnara is model-agnostic: developers bring their own keys for OpenAI, Anthropic, OpenRouter, Amazon Bedrock, LiteLLM or Ollama, or point it at any compatible endpoint, including models they host themselves. Built-in tools cover shell commands, file operations, web search and fetch, and asking a human a question, alongside support for MCP servers and custom tools that run inside a customer's own systems. Agents can load packaged bundles of instructions and files, known as skills, on demand; share persistent memory stores with other agents under read or read-write access; and spawn subagents that run with fresh context and can themselves be inspected.

Machines and durability

An agent can work across several machines at once and move between them — the README's example is a set of GPU servers — and if a user's laptop goes offline, the agent can continue in a sandbox and sync its work back with git. Sandboxes are available from eleven providers, including Daytona, E2B, Modal and Unikraft. Cron-style schedules with timezone support can message an existing agent, launch one from a profile, or open a new Slack or Discord thread.

Integrations and permissions

Mentioning the bot in Slack or Discord starts an agent in a thread where it can reply, ask questions and request approvals. On GitHub, an agent can be triggered when a pull request opens or when someone mentions it, then read the diff, comment and submit reviews. Every event can be consumed as a stream or a webhook, new messages can be sent while the agent works, and per-tool permission settings determine whether a tool runs immediately or must ask first.

Self-hosting and cloud

Self-hosting requires Docker with Compose: after cloning the repository and starting the stack, users sign up on a local instance, and the verification link is printed in the container logs so no email provider is needed for development. A fresh instance ships with no models or machines configured, so operators add a model provider, register a machine and share both with their project. The README warns that the local defaults are deliberately insecure and points to separate guides for real deployments. On Omnara Cloud, a first project comes pre-configured with a model provider called omnara-openrouter and a pool of managed sandboxes, so the quickstart configuration runs without further setup.

Why it matters

Managed agent APIs are convenient but tie teams to a single vendor's models, pricing and infrastructure. Omnara's bet is that the same capabilities — durable state, approvals, streaming and integrations — can be delivered as open source running on hardware you control, with whichever model you choose. The Postgres-backed event log in particular brings a degree of fault tolerance usually associated with hosted services. The trade-off is operational: running it yourself means operating Postgres, Redis, blob storage and machine daemons, and hardening a stack whose development defaults are intentionally loose. For teams wary of letting agents loose on real machines, an Apache 2.0 option they can audit and host internally is a meaningful alternative to committing to one cloud provider.