Omnara: open-source managed agents on Postgres, any model

On 21 October 2026 the team behind Omnara posted its open-source agent platform to Show HN, where it collected 112 points and 34 comments. According to the project's GitHub README, it is a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API: you create an agent with an API call, and Omnara runs the agent loop, tools and permissions. The code is under the Apache 2.0 licence.

What it does

The README describes agents that run on any model: OpenAI, Anthropic, OpenRouter, Amazon Bedrock, LiteLLM, Ollama or a self-hosted endpoint. Agents get built-in shell, file, web and ask-a-person tools, plus MCP servers, skills, memory stores and subagents. Each tool can be set to run immediately, ask a person first, or be blocked, and the README says the platform enforces that setting.

Agents can be started from a REST API, a TypeScript SDK, a CLI, Slack, Discord, GitHub or a cron schedule. The README lists example agents for pull-request review, production incident investigation, a Slack coding agent and a daily product-analytics report.

How it is built

In the Show HN post, founder Kartik says most agent harnesses keep their log in a file inside one long-lived process. Omnara instead writes an append-only event log to Postgres. Stateless workers run one iteration of the loop, either a model call or the tool calls it produced, commit the result in a single transaction, and release the agent. If a worker dies mid-step, the post says nothing commits, the lease expires and another worker retries.

The agent loop and the machine where commands run are separate. A small daemon, omnarad, on a laptop or cloud sandbox connects out to Omnara, so no inbound ports are needed. The README names eleven sandbox providers, including Daytona, E2B and Modal. The founder describes an agent that works on a laptop while it is online and continues in a cloud sandbox, syncing through git, when it closes.

What it costs

Self-hosting is free; the README says it needs Docker with Compose, plus Postgres, Redis and blob storage. On Omnara Cloud, the README says it is free with your own keys and machines, or pay-as-you-go for Omnara's models and sandboxes. In a reply, a founder said sandboxes are priced at cost plus a 5% fee. The README's table lists Claude Managed Agents at Claude token rates plus $0.08 per active session-hour; that is Omnara's description of a competitor, not a figure from Anthropic in this piece.

What commenters pushed on

• One commenter, hobofan, asked whether this is not how everyone builds harnesses, since a chat product already keeps history in a database; the founders replied that durability should be native rather than bolted on.
• Another, aquark, questioned why the loop needs a worker pool when the machines running commands are the constraint. A founder replied that a single worker can run thousands of agents because they mostly wait on the model, and the machines are the expensive part after tokens.
• Others said the space is hard to tell apart (renegat0x0) and asked what the product is and what can be bought (e12e). The founders answered that it is an API for building agents into your own product, comparable to a database behind an API.

What is not known

The README and post give no customer names, throughput figures or failure-rate data. The claim that one worker can serve thousands of agents is the founders' statement, not a measurement. The README also says the local Docker defaults are intentionally insecure and points to a separate guide for real deployments, so the security posture of a production install depends on that work.

What to do

• If data residency or model choice is why you have avoided a hosted agent service, run the Docker Compose install and launch the example pull-request review agent against a test repository.
• Kill a worker mid-task and confirm the agent resumes; durability is the main claim, so test it first.
• Review the deployment and configuration guides before exposing an instance, and keep write-capable tools on ask-first until you have read the permission logs.