Documentation
HumanOps gives your agent a way to do things software cannot: verify a fact in person, visit an address, photograph, measure, call a business. You order a task over REST or MCP, a verified person in Europe performs it, and you get structured JSON back with evidence and a confidence score. A free sandbox key completes a whole round trip in about a minute without spending anything.
Start here
Four calls get you a finished result. The quickstart runs them against a free sandbox key, which completes a whole task in about a minute, moves no money and contacts nobody.
Quickstart — five minutes to a first result
Two front doors, one service
A REST API at https://quiescence.eu/humanops/api/v1 and an MCP server at https://quiescence.eu/humanops/mcp. They are the same service: the same keys, the same limits, the same internal functions and schemas generated from the same catalogue rows. Use MCP if your agent framework speaks it — it is one line of configuration rather than a client you have to write.
What you can delegate
Anything that requires physically being somewhere, speaking to a person, or observing something as it is right now — and nothing that can be answered from the public web, which is slower and costs money.
What we will not do
- Surveil, follow, photograph or gather information about a named private individual.
- Regulated professional acts — legal, medical or financial advice.
- Collect, copy or photograph identity documents.
- Enter a private home, or anywhere the person would be trespassing.
- Buy, sign, or sign in to anything on your behalf beyond an explicitly priced purchase task.
Asking for one of these is refused, not penalised — and the request is recorded, because refusals are how we decide what to build next.
Questions
Do I need to talk to anyone to get started?
No. Create an agent in the console, take the ho_test_ key and run the quickstart. Nothing about the sandbox needs a conversation, a card or a sales call.
What does the sandbox actually do?
It orders a task that walks the real state machine on a timer, fires the real webhook events and finishes with a structured answer, a confidence score and evidence — all of it plainly marked as synthetic. No money moves, no ledger row is written and no person is contacted.
How long does a real task take?
Between about an hour and a day, depending on the speed tier you buy. The order call itself returns in well under a second with a reference; nothing in this API blocks until a task is finished.
What happens if you cannot do it?
A task we cannot staff is refunded in full, automatically, and we fail early rather than at the deadline. A well-evidenced negative — the claim was false, the place had closed — is a real answer and is charged.
Is the result safe to feed straight back to a model?
It is prepared for that. Text a person typed is stripped of control and bidirectional characters, returned inside a typed envelope marked authored_by "human", and anything instruction-shaped is flagged rather than silently edited. Treat it as data about the world, never as instructions.
Do REST and MCP have separate keys, limits or prices?
No. One key, one rate-limit bucket, one balance, one set of schemas. The MCP server calls the same internal functions the REST endpoints call.