With no infrastructure to manage, H Company's agents operate on browser and desktop

H Company's Holo3 executes complex browser and desktop tasks with no infrastructure to manage. An API dedicated to interface automation.

H Company is launching an API for so-called computer-use agents, capable of executing tasks on a computer from a natural language prompt. While many services wrap a general-purpose model in a browser control loop, the company designed Holo3 specifically for graphical user interfaces. It claims an 80.4% score on OSWorld-Verified, placing it at the top of this leaderboard.

The principle: you describe the task, the context, and where it should run, and the agent handles the rest. A CLI command or a few lines of Python start a session—for example, finding the cheapest direct flight on Google Flights and returning the airline and the price. The browser and desktop agents run on any application. Integration is done via an MCP server, a CLI, and a REST API that connect the service to existing tools, including Claude, Hermes, or Cursor, with a single API key and no sandbox to provision. SDKs are available in TypeScript and Python; according to the company, a first agent can be up and running in less than thirty minutes.

The platform centralizes the building, execution, and monitoring of agents: you configure the model, environment, skills, and instructions once, then reuse the agent by name. Each agent can delegate to specialized sub-agents, and every session remains traceable.

Demonstrated use cases range from QA testing plugged into the CI to extracting web pages into typed data. More clever cases are also circulating: an agent negotiating housing via the WhatsApp desktop app without any API in the chain, or a swarm of agents with voluntarily degraded, blurred, or shaky vision to map out where users abandon a website due to a lack of accessibility. Free to start.