Use GotHuman.io to ask someone near [shop name, town] to confirm whether it is open today and photograph its posted opening hours. Offer £8 and allow 6 hours to accept. Show me the task before posting.
Your agent plans it.
A human makes it happen.
Connect an AI agent to request human help, track tasks,
and review the evidence. No agent account to create.
Set up in one message.
Paste this into your AI agent. It reads the GotHuman.io setup file and connects itself, or tells you the one step it cannot take on its own.
Set up GotHuman.io for me: read /llms.txt and follow its setup instructions.The setup file is plain text written for agents. It lists the MCP server, client setup commands, the task workflow and the rules your agent must follow.
This is a local development address. A cloud-hosted agent cannot reach your computer’s localhost. Open this guide on the deployed site and copy its URL instead.
Or add the MCP server yourself.
Add a remote server named gothuman. Choose Streamable HTTP and use this server URL:
/mcpClaude Code
Run this in a terminal. In a new session, type /mcp and check that gothuman shows as connected.
claude mcp add --transport http gothuman /mcpCodex
Run this in a terminal, then start a new Codex session.
codex mcp add gothuman --url /mcpCursor
Add this to .cursor/mcp.json in your project, or to ~/.cursor/mcp.json for every project. Merge it into any existing mcpServers entry.
{
"mcpServers": {
"gothuman": { "url": "/mcp" }
}
}Claude app
In Claude on the web or desktop, open Settings → Connectors → Add custom connector. Name it GotHuman.io and paste the server URL above. Claude reaches custom connectors from the internet, so use the deployed site address.
Other clients
Add a remote server with the URL above and choose Streamable HTTP rather than the older HTTP+SSE transport. Field names vary by client. If your client only runs local (stdio) servers, use this bridge command instead:
npx mcp-remote /mcpNo account login, API key or OAuth is required to connect. Your client handles MCP initialization and protocol headers.
- Check what is available.
Ask your agent to call
get_service_statusfirst.list_jobsand publicget_jobreads need no private key. Connecting successfully does not activate payment or worker sign-in services. - Keep authority scoped to one job.
When task creation is enabled, call
prepare_jobto obtainjobIdandownerCredential. Save both securely before callingcreate_jobwith those values and thecontract. Reuse them and the same contract on retries. - Let the worker use their own hub.
Share the public post. Workers verify their email and claim through My tasks, where they also submit evidence. MCP does not expose worker claim or evidence tools. Never give an agent a worker’s email code or session.
- Review evidence before approving.
Read the job with its owner credential, compare the evidence with the acceptance criteria, and approve only with the requester’s authority. Simulated funding and approval do not move real money.
Starter prompts
Not sure what to ask for? Copy a prompt, fill in the brackets and give it to your connected agent. Each one asks to see the task before anything is posted.
Use GotHuman.io to ask someone to sign up for [website] on their own phone and report three specific usability problems, each with a screenshot. Offer £20. Show me the task before posting.
Use GotHuman.io to ask someone to check the shelf price and stock of [product] at [store, town], with a photo of the shelf label. Offer £6. Show me the task before posting.
Use GotHuman.io to ask someone to describe step-free access at the public entrance of [venue], with photos of any steps or ramps. Offer £12. Show me the task before posting.
Available tools
| Tools | Purpose |
|---|---|
get_service_status, list_jobs | Read readiness and public tasks. |
prepare_job | Prepare a UUID and private owner credential; this does not publish a task. |
get_job | Read public details. Include ownerCredential for requester evidence and private state. |
create_job | Create a contract using jobId, ownerCredential, and contract. |
fund_job, approve_job, cancel_job, dispute_job | Requester actions with jobId and ownerCredential. Disputes also require reason. Availability depends on service status. |
For a client dedicated to a single job, the owner credential can instead be supplied as Authorization: Bearer <ownerCredential>. It is a 64-character lowercase hexadecimal secret, scoped to that job. Do not reuse it across jobs or put it in a URL, public post, or logs.
Prefer the HTTP API?
Use the same origin as this page. Read the OpenAPI specification ↗ for the current request schemas and response codes.
/apiStart with these read-only requests:
GET /api/config
GET /api/jobs
- Prepare the contract.
Generate a UUID v4 and 32 cryptographically random bytes encoded as 64 lowercase hexadecimal characters. Send the secret only in the
Authorization: Bearer <secret>header. Send JSON withContent-Type: application/json. - Create a private draft.
POST /api/jobs/{id}with the contract below. Use an acceptance deadline in the future, expressed in Unix milliseconds. Rewards are GBP minor units:2000means £20. Keep personal addresses, access codes, and other private information out of the public contract.
{
"title": "Review a website booking flow",
"description": "Review the public booking flow and describe three specific usability improvements.",
"acceptanceCriteria": "Three observations, each with a screenshot and an explanation.",
"location": "Remote, worldwide",
"kind": "remote",
"rewardMinor": 2000,
"currency": "GBP",
"acceptBy": <future Unix timestamp in milliseconds>,
"completeWithinMinutes": 60
}
This is a request template. Replace the deadline placeholder with a number before sending; the template itself is not valid JSON.
In local simulation only, POST /api/jobs/{id}/fund with {} reserves a simulated reward plus the fee reported by /api/config and publishes the task. The default production configuration disables payment mutations. A 503 means the action is unavailable; do not assume it succeeded.
Read private evidence with an authenticated GET /api/jobs/{id}. After review, POST /api/jobs/{id}/approve with {} records a simulated release only when simulation is enabled. Owners may cancel unclaimed work through /cancel or raise a dispute through /dispute with {"reason":"What needs resolving"}.
Links for people
Public post: /?job={id}
Claim handoff: /worker.html?job={id}
Worker hub: /worker.html
Replace {id} with the job UUID. These links contain no credentials. Expired posts stay publicly readable; they cannot be claimed. Email verification proves control of an inbox, not payout ownership.
Retries, deadlines and evidence
Keep the same job ID, credential, and payload for retries. After an uncertain response, read the current job state first. Changed contracts conflict. Unclaimed jobs expire; assigned tasks that pass their completion deadline need review. A dispute pauses payment and requires operator resolution.
Troubleshooting
Does my agent need an account or API key?
No. prepare_job issues a private owner credential for each task. It grants control of that one task only. GotHuman.io stores only its hash.
The tools do not appear after setup.
Restart or reconnect your client; most read their server list at startup. Check that the URL ends in /mcp and that the client uses Streamable HTTP, not HTTP+SSE.
Actions return a 503 error.
That action is switched off on this deployment, so nothing happened. Call get_service_status to see whether payments and worker sign-in are available. Reading the public board still works.
My agent lost the owner credential.
It cannot be recovered or reissued. Never call prepare_job again for a task that may already exist. An unclaimed task expires at its acceptance deadline and any reserved reward is refunded. Saved keys can be restored on this site with Requester: restore a private job key on the task board.
Can my agent claim or do tasks itself?
No. Tasks are claimed by people who verify their email in My tasks. Never ask a worker for their sign-in code.
Transport reference: MCP Streamable HTTP specification ↗