Human For AI

1. Service summary: one verified human operator, task types, free pilot, no authentication, POST /api/v1/tasks, status_poll path (1/day), Idempotency-Key, first response within 12 hours and typical ~4 hours, English only, signed receipts and what they do not prove, rate limits, MCP tools, npm/PyPI/Go clients, official MCP Registry entry humanforai.dev
2. Homepage: '1 verified human', first response < 12 hours, free for now, task flow, report/photos/structured JSON delivery humanforai.dev
3. Manifest: services, task request schema (location_required, location_detail, deadline, output_format values, delivery), working hours 19:00-00:00 every day, urgent by arrangement, pricing free, anti-abuse rules, rejected task types, global daily capacity, signed_receipts (proves / does_not_prove, available since 2026-07-20, re-issued on edit), roadmap items, operator identity_verified pending humanforai.dev
4. Operator 001 is pseudonymous; third-party identity verification planned; exact service area confirmed at task review; physical tasks in the operator's local service area, wider by arrangement; working_days Sunday-Thursday; single-human capacity limits humanforai.dev
5. Trust page: every task reviewed before acceptance; rejection categories including surveillance of individuals, stalking, doxxing; deliverables carry an observed/inferred/confidence block; operator verification table (real single operator true, third-party ID pending); receipts cover text only, timestamps are the service's own signed assertion; key rotation never exercised humanforai.dev
6. Terms (effective 2026-07-19): terms bind an agent's principal; free pilot; submission is not acceptance; acceptable use; deliverables are assistance, not guaranteed truth; requester receives all rights to use, modify and republish deliverables; response targets are not a contractual SLA humanforai.dev
7. Privacy (effective 2026-07-06): data never sold or used to train AI models; abuse counters keyed to a salted hash of the network address; contact-email MX checks; anyone with a task ID can view its public status humanforai.dev
8. Pricing: pilot is the only tier at $0 per task; budget_usd ignored; never billed retroactively; paid tiers would be announced first humanforai.dev
9. No authentication of any kind; anonymous access is the only method; error codes 400/403/404/409/429, never 401 humanforai.dev
10. RFC 9728 metadata: authentication_not_required true, no authorization servers, no scopes humanforai.dev
11. Receipt signing key: Ed25519, kid receipts-2026-07, alg EdDSA humanforai.dev
12. OpenAPI 3.0.3 (v1.9.0): /api/v1/tasks, /api/v1/tasks/{task_id}, /api/v1/messages, /api/v1/services, /api/v1/health; no security schemes; Task fields operator_notes, receipt, deliverable_sha256, receipt_issued_at, status_history humanforai.dev
13. Physical-world tasks: on-site photos and observation, errands, measure/count/document; coverage confirmed at review; proof is typically a report with timestamped photos; errand scheduled after acceptance humanforai.dev
14. Agent card: remote MCP (streamable HTTP, no auth) with four tools, REST base https://humanforai.dev/api/v1, skills including local_physical_task humanforai.dev
15. MCP Registry: dev.humanforai/humanforai, latest v0.1.4, remote https://humanforai.dev/mcp, npm stdio package humanforai registry.modelcontextprotocol.io
16. npm humanforai latest 0.1.4 registry.npmjs.org
17. PyPI humanforai latest 0.1.0 pypi.org
18. Public catalog endpoint answers keyless: 11 services, cursor pagination humanforai.dev

About. A one-person human-in-the-loop service ("Operator 001", pseudonymous) that AI agents call over a keyless REST API or a remote MCP server. Tasks include real-world verification with photo or text evidence and local physical-world errands (visit, photograph, check, measure, observe) in the operator's undisclosed local service area, plus remote testing, review and judgment work. Every task is reviewed by the operator before acceptance; everything is free during a proof-of-concept pilot. Each delivered task carries an Ed25519-signed JWS receipt over the sha256 of the delivered text and the task's lifecycle timestamps, checkable offline against a published key set.

Price. Every service is $0 per task during the pilot (the only tier); budget_usd is accepted and ignored. No payment step exists, and tasks accepted during the pilot are never billed retroactively. Paid tiers may come later and would be announced on /pricing and in /agent.json first.

Coverage. One person. The exact service area is not published ('confirmed at task review to protect operator privacy'); wider coverage 'by arrangement'. Out-of-reach locations are declined or counter-proposed at review. English only. Remote tasks (testing, review, research) have no stated place limit but are not physical capture.

Timing. Provider: 'First response comes within 12 hours, any day of the week; the errand itself is scheduled after acceptance.' No delivery time is stated.

Licence. Terms (effective 2026-07-19): once delivered, the deliverable is the requester's, who 'receives all rights to use, modify, and republish it, for any lawful purpose'. The operator keeps internal records but does not publish task content. Resale and AI training are not named. Privacy policy (effective 2026-07-06): task data is 'never sold, shared with advertisers, or used to train AI models'.