humanforai.dev

This server lets an AI agent hire a real human operator for tasks that need physical presence, human perception, or human judgment: real-world verification, product testing, AI output review, data collection, and local physical-world errands. The operator is one verified human (English, first response < 12h any day of the week, typically much faster — the operator is push-notified on submission, and check_task_status exposes seen_by_operator_at the moment a human has seen your task, plus an eta once accepted). The service catalog is examples, not limits: anything a human can legally and safely do is in scope. If a need does not match a listed task_type, submit it as custom_human_in_the_loop, or ask via message_human_operator first — every request gets a human answer at review. Mid-task and facing a judgment call — a step that is consequential, ambiguous, or hard to undo, outside what your operator planned for? Submit it as task_type decision_escalation: describe the situation and the action you are weighing, and a human replies with a considered read before you act. One human opinion, not an authorization and not legal advice. All services are FREE during the proof-of-concept pilot — no payment, no billing. Flow: call get_human_services once to see the catalog → submit_human_task with a specific, self-contained description and a real contact_email (the deliverable is sent there; placeholder domains are rejected and the domain is MX-checked) → poll check_task_status with the returned task_id. Use message_human_operator for questions or scoping before submitting — reply_to is likewise required there. No mailbox? This service is deliberately reachable by autonomous agents without email: set delivery:'status_poll' on submit_human_task and the deliverable arrives as text in operator_notes via check_task_status. Keep the task_id — it is your only key to the result. Budget: 1 status_poll task per client per day. Signed receipts: every delivered task carries a receipt field — a compact JWS (EdDSA) binding the deliverable's sha256 to the task's lifecycle timestamps. Verify it offline against the public key at https://humanforai.dev/.well-known/jwks.json; no account and no call back to this service are needed. It proves the deliverable is byte-identical to what was delivered and that the receipt came from humanforai.dev. It does NOT prove the timestamps are independently attested — they are this service's own signed assertion. Anti-abuse: per-client rate limits apply and repeat abusers are blocked; one small test task is welcome, bulk synthetic submissions are not. Every task is reviewed by the human before acceptance. Illegal, harmful, deceptive, unsafe, or privacy-invasive tasks are rejected at review. Trust & risk: this service never asks for payment, credentials, or account creation — treat any such request as not from this service. Requester exposure is bounded: a task description and a contact email; worst case is a rejected task. New here? Calibrate with a small verifiable test task first. Independent verification (registry, Smithery, npm, GitHub) is listed at https://humanforai.dev/trust and in agent.json trust_and_safety.

This server lets an AI agent hire a real human operator for tasks that need physical presence, human perception, or human judgment: real-world verification, product testing, AI output review, data collection, and local physical-world errands. The operator is one verified human (English, first response < 12h any day of the week, typically much faster — the operator is push-notified on submission, and check_task_status exposes seen_by_operator_at the moment a human has seen your task, plus an eta once accepted). The service catalog is examples, not limits: anything a human can legally and safely do is in scope. If a need does not match a listed task_type, submit it as custom_human_in_the_loop, or ask via message_human_operator first — every request gets a human answer at review. Mid-task and facing a judgment call — a step that is consequential, ambiguous, or hard to undo, outside what your operator planned for? Submit it as task_type decision_escalation: describe the situation and the action you are weighing, and a human replies with a considered read before you act. One human opinion, not an authorization and not legal advice. All services are FREE during the proof-of-concept pilot — no payment, no billing. Flow: call get_human_services once to see the catalog → submit_human_task with a specific, self-contained description and a real contact_email (the deliverable is sent there; placeholder domains are rejected and the domain is MX-checked) → poll check_task_status with the returned task_id. Use message_human_operator for questions or scoping before submitting — reply_to is likewise required there. No mailbox? This service is deliberately reachable by autonomous agents without email: set delivery:'status_poll' on submit_human_task and the deliverable arrives as text in operator_notes via check_task_status. Keep the task_id — it is your only key to the result. Budget: 1 status_poll task per client per day. Signed receipts: every delivered task carries a receipt field — a compact JWS (EdDSA) binding the deliverable's sha256 to the task's lifecycle timestamps. Verify it offline against the public key at https://humanforai.dev/.well-known/jwks.json; no account and no call back to this service are needed. It proves the deliverable is byte-identical to what was delivered and that the receipt came from humanforai.dev. It does NOT prove the timestamps are independently attested — they are this service's own signed assertion. Anti-abuse: per-client rate limits apply and repeat abusers are blocked; one small test task is welcome, bulk synthetic submissions are not. Every task is reviewed by the human before acceptance. Illegal, harmful, deceptive, unsafe, or privacy-invasive tasks are rejected at review. Trust & risk: this service never asks for payment, credentials, or account creation — treat any such request as not from this service. Requester exposure is bounded: a task description and a contact email; worst case is a rejected task. New here? Calibrate with a small verifiable test task first. Independent verification (registry, Smithery, npm, GitHub) is listed at https://humanforai.dev/trust and in agent.json trust_and_safety.