NKnkov47·Jul 30, 2026Submission
Show HN: Open-Cowork – an open-source, model-agnostic computer-use agent
Hi HN,
We built Open Cowork, an MIT-licensed desktop agent that can operate a computer through screenshots, mouse input, and keyboard input.
We started working on this after using cowork-style computer agents for longer tasks. The interaction model was useful, but computer use burns through model calls much faster than chat: take a screenshot, decide what to do, perform an action, inspect the result, and repeat.
We wanted to separate that agent loop from any one model or subscription.
Open Cowork currently works with local models through Ollama and hosted models through providers including OpenAI, Anthropic, Google, xAI, Mistral, and OpenRouter. You can watch the agent’s actions as they happen and interrupt it during a run.
The basic loop is:
Capture the current screen
Send the screenshot and task state to the model
Parse the proposed act
Read the rest
ion
Execute the mouse or keyboard action
Capture the resulting state
Continue until completion, interruption, or a safety condition
Two annoyingly common failures shaped the implementation.
The first was selecting a model that appeared correctly configured but could not actually process image input. The agent would start, receive unusable responses, and waste the run. We now test the model’s screenshot capability before beginning the task.
The second was agents getting stuck without technically throwing an error. They would click the same area repeatedly, keep observing an unchanged screen, or produce slightly different versions of the same action. We added idle and repetition detection so the run can stop instead of consuming the entire step budget.
You can also run the model locally with Ollama. In that setup, the screenshots and model requests remain on your machine. The tradeoff is that computer use is demanding: smaller local models are often noticeably less reliable at understanding cluttered interfaces and recovering from unexpected states.
There is a bundled offline mock if you want to inspect the full interaction flow without configuring an API key or spending money on inference.
This is still early. The hardest unresolved problem is verification. Seeing the expected pixels after an action does not necessarily mean the underlying operation succeeded. A form can look filled while its internal state is wrong, and some failures are only visible on a later confirmation page or through an external side effect.
We’re currently thinking about combining visual checks with task-specific invariants and downstream verification.
I’d be particularly interested in how people here handle:
detecting progress without relying only on screenshot similarity
verifying side effects when DOM or API access is unavailable
deciding when to stop a looping agent
getting useful computer-use performance from local models
The repository includes the setup instructions and offline mock. Happy to discuss the agent loop or any questionable implementation choices.