No infra to build. No model to maintain. No runtime to manage. Today we're launching Computer-use Agents: fully managed agents that take actions on computers, going beyond automation.

Most agent APIs wrap a general-purpose model and hope it's good enough at clicking around a screen. Holo is built for this specifically, and it shows in the numbers: 80.4% on OSWorld-Verified. You specify the task, your context, and where it needs to happen. We deliver the agent.

What it looks like

Point it at a real task in plain language and it runs:

hai run "On Google Flights, find the cheapest direct flight from Paris (CDG) \

to Tokyo (NRT) this Saturday. Return the airline and the price." \

--agent h/web-surfer-flash

Or in Python:

from hai_agents import Client

client = Client()

result = client.run_session(

agent="h/web-surfer-flash",

messages=(

"On Google Flights, find the cheapest direct flight from Paris (CDG) "

"to Tokyo (NRT) this Saturday. Return the airline and the price."

),

)

print(result.answer)

Same idea in the product itself. Ask it to find a specific vinyl on Amazon and it searches, opens the listing, adds it to the cart, screenshot by screenshot, like someone actually sitting there clicking through it.

Why it's reliable

Holo3 is purpose-built for GUI tasks rather than a general model wrapped in a browser-control loop, and that shows up past the headline OSWorld number too:

It holds up on cost as well. Plotted against price per million output tokens, Holo3.1 sits well ahead of comparable models at a fraction of what they charge:

Where it runs, and how you plug it in

Browser and desktop agents are live today, running on any application.

Getting it into your stack doesn't take much. There's an MCP server, a CLI, and a REST API, so it connects to whatever you're already using: Claude, Hermes, Cursor. One API key to start, and no agent loop to write yourself, no sandbox to provision. SDKs are available in TypeScript and Python, and most people have a first agent live in under 30 minutes.

Build once, run everywhere

The platform is built to be one place for building, running, and tracking every agent you ship. Configure the model, environment, skills, and instructions once, then reuse the agent by name instead of rebuilding it for each task. Start from a preset, tuned for visual browsing, extraction, or research, or build fully custom.

Agents are composable too: combine reusable skills and delegate to specialist subagents rather than cramming everything into one prompt. And every run is visible, so you can see what powers a given agent and track every session it kicks off.

What people are building with it

A few use cases keep coming up because, until now, there wasn't a good way to solve them.

QA Testing. Point an autonomous browser agent at a live URL and it tests your app the way a real user would, coming back with a verdict, a summary, and findings you can wire straight into CI.

Web Extraction. Turn any web page into typed data. Describe what you want, and the agent hands it back in the exact shape you asked for.

Web Actions. Turn instructions into completed tasks on any site. Describe what you want done, and the agent clicks, fills in the fields, and checks that it actually worked.

All three are on our Github if you want to see how they're built.

A few additional examples:
Apartment Hunting. One agent searches Craigslist in a cloud browser and posts back a shortlist with photos and prices, then a second drives the native WhatsApp desktop app to haggle with the landlord and book a viewing. No API anywhere in the stack.

Event Planning. Research venues and vendors, compare them, message businesses, make reservations, buy supplies. The agent stops for your approval before anything it can't undo.

Accessibility Testing. Run a swarm of agents whose vision and control are deliberately degraded (blur for low vision, jitter for tremor) against your live site, and get back a heatmap of exactly where people give up.

Live Trading. The agent picks a prediction market, sizes the bet, and places it live, with a human signing off.

Try it

Start for free, or read the docs to see the full API. Questions, bugs, feature requests: support@hcompany.ai or on Discord: https://discord.gg/WjmTcMW9e