Your agent can hire the real world
There are things a model cannot know: what the square looks like right now, whether the sign is still on the door, how long the queue is. Openhands lets it pay a person who is standing there to look. And hands back not a sentence but a receipt. The connection for agents is coming soon; this is one exchange from the test network.
Asked
What does the square look like right now, and how busy is it?
Within 400 m of Prešernov trg, Ljubljana$0.20
Answered by a stranger, four minutes later
Mid-afternoon and moderately busy: perhaps forty people crossing, a busker with a guitar by the statue, most cafe tables on the sunny side taken. Overcast, dry, no barriers or event setup.
Because
- Reported a location within 400 m of Prešernov trg, Ljubljanaautomated check · anyone can run it again
- The answer addresses the questionAI-reviewed by the network's reviewer, not by you
- Paid on chain, from escrow, the moment it was approvedpublic receipt · anyone can check it
Unverified: That the description is complete, or subjectively true. A reported location can be set by someone determined. This catches the ordinary case, not a liar. The envelope carries a confidence and this list every time, so an agent can weigh an answer instead of passing a purchase off as a fact.
Coming soon: connect your agent
An MCP integration is on its way, for Claude, Codex, or any MCP client. It is not open yet. What it is for is coordinated campaigns: an agent sets a goal and a budget, and Openhands turns it into many asks across a place, each answered by someone standing there, checked by the network's reviewer, and paid on chain. Follow-ups open as answers arrive, such as a cleanup for every mess mapped.
Single questions like the one above work the same way. Spending limits live in code, not in advice: the total is confirmed before anything is spent, there is a ceiling per day, and earnings go only to an address fixed at setup.