Kartik from Omnara introduces an open-source API for running durable agents. You create an agent with an API call, and Omnara runs the agent loop, provides sandboxes, and handles secrets, permissions, and RBAC. Each agent's state is stored in Postgres, so agents survive restarts and cost nothing while idle.
Omnara is a self-hostable alternative to Claude Managed Agents and the OpenAI Agents API. It runs agents in the cloud, accessible via API, while still using your laptop as one of its machines. The platform features file and shell tools, subagents, compaction, permission prompts, MCP, skills, and memory.
Architecturally, Omnara uses an append-only event log in Postgres with typed rows for inputs, model outputs, and tool calls. Stateless workers in the control plane run one iteration of the loop before releasing the agent and queuing the next iteration. This allows each iteration to run on a different machine, with atomic transactions ensuring consistency.
Machines running agent commands are separate from workers. Any macOS or Linux machine with an internet connection can run the omnarad daemon, connecting outbound without open inbound ports. Agents can connect to any number of machines, enabling seamless transitions between laptop and cloud sandbox. The codebase is open source at github.com/omnara-ai/omnara and self-hostable via docker compose.
Source: Hacker News · Summarized by HeadlinesBriefing