omega: an AI agent platform that runs on your machine

Every AI agent platform wants your data in someone else's cloud. Your prompts, your files, your API workflows - they run on hardware you don't control, under terms you didn't read, on a subscription you keep paying. And when you build something real with an agent, you build it inside someone else's product. They can change the price, the model, or the whole roadmap. Your setup lives or dies with their roadmap.

I covered the model side of this in my first post: run the LLM locally. But a model is just an engine. What's missing is the rest of the car - the workspace where an agent actually does things.

That's omega. A local-first AI agent platform: one Go binary, no accounts, no telemetry, binds to 127.0.0.1 by default. You run it on your machine - a Raspberry Pi is enough - and the agent works for you with your models. Ollama locally, or any OpenAI- or Anthropic-compatible endpoint if you want to point it elsewhere. Your call, per provider.

It ships as two frontends that share one brain: a web UI and a terminal UI. Same sessions, same data, whichever you're in the mood for.

• Tools that do real work: shell, file editing, persistent memory, web search and fetch, image input, an MCP bridge. The agent doesn't just talk about your project - it works in it.
• Sessions that branch: retry a response and omega creates a sibling variant instead of wrecking your transcript. Flip between paths, fork from any message, keep the pick on reload. This sounds small. After a week you can't live without it.
• Skills: teach it procedures once - in plain markdown files - and it repeats them correctly every time.
• Cron: scheduled prompts on a timer. Report every morning, backup check every night, whatever you automate.
• Kanban: a task board where cards can be built by agent runs, not just by you.
• Delegation: hand a subtask to a subagent and get the result back in the main conversation.

The web UI is what you'd want from a workspace, not a chat demo: a VS Code-style shell where panes dock, drag and resize, and carry their own tabs. Sessions, files, settings, the board - each in its own pane, on desktop and mobile widths.

Local-first is the architecture, not the marketing. The server binds to loopback unless you explicitly open it. URL fetching refuses internal addresses. No account system, no telemetry, no "anonymous usage stats". Nothing to leak, because nothing leaves.

One binary from GitHub - EndoTheDev/omega. Six platforms, including a Raspberry Pi near you. omega serve and the web UI is up. Updates are built in - omega update fetches the next stable, --nightly if you like living on the train.

It's developed in the open. If local agents are your thing, the roadmap is public and the issues board is open.

The models got local. Now the agent is too. One binary, your hardware, your data - and it works offline, because it never needed the cloud in the first place.