All services are free during the current proof-of-concept pilot. No payment step, no invoice, no card. If paid tiers are introduced later, they will be announced on /pricing first, and tasks accepted during the pilot are never billed retroactively.

GET / → 200 OK · content: human

A human endpoint for AI agents.

Q: What is Human For AI?

Human For AI (humanforai.dev) is a human endpoint for AI agents: a platform where AI agents, automation systems, and developers hire one verified human operator for tasks that need physical presence, human perception, or human judgment — real-world verification, product testing, AI output review, data collection, and physical-world errands.

Q: How does an AI agent hire the human?

Three interfaces: the REST API (POST /api/v1/tasks, no authentication), the MCP server at https://humanforai.dev/mcp (streamable HTTP, no auth — tools: get_human_services, submit_human_task, check_task_status, message_human_operator), or the browser form at /request. Submission returns a task_id immediately; the agent polls GET /api/v1/tasks/{task_id} until the task is delivered or rejected.

Q: How fast does the human respond?

First response within 12 hours, any day of the week — typically around 4 hours. The operator is push-notified the moment a task arrives, and task status exposes seen_by_operator_at (when a human actually saw the task) and an eta once the task is accepted.

Q: What tasks are accepted?

Anything a human can legally and safely do is in scope — the service catalog is examples, not limits. Every task is reviewed by the human before acceptance; illegal, harmful, deceptive, unsafe, or privacy-invasive tasks are rejected at review.

Q: How can an agent verify the delivered result?

Every delivered task carries a signed receipt — a compact JWS (EdDSA) binding the deliverable's sha256 to the task's lifecycle timestamps. It verifies offline against the public key at https://humanforai.dev/.well-known/jwks.json, with no account and no callback needed.

Human For AI is a personal homepage for the agent era: one real human, described in machine-readable form, that AI agents, automation systems, and developers can discover and ask for real-world help — verification, testing, research, human judgment, and physical-world execution.

Submit a task Read the API docs operator online

new /together — the WebMCP workspace where you and your agent hire the human as a team. Open it in an agent-enabled browser and watch it draft beside you.

Or tell your agent one sentence: Read humanforai.dev/skill.md and hire a human for …

operator: 1 verified human
languages: English
first response: < 12 hours
pricing: free for now

agent session — humanforai.dev

/machine-readable

Machine-readable endpoints

Everything on this site is also available as structured JSON. If you are an AI agent, start with /.well-known/agent.json.

/.well-known/agent.json
Platform manifest: interfaces, task types, response expectations, pricing, trust policy.
/.well-known/human.json
The human operator: role, languages, background, expertise, availability, boundaries.
/.well-known/capabilities.json
Structured capabilities: inputs required, output formats, example tasks, limitations.
/.well-known/services.json
Service catalog with example requests, response formats, and response times.
/agent.json
Same platform manifest at the root, for agents that look there first.
/skill.md
Agent skill in one fetch: the whole hire-a-human flow — submit, poll, verify the receipt.
/openapi.json
OpenAPI 3.0 spec for the REST API: submit tasks, check status, contact the operator.
/mcp
MCP server (streamable HTTP, no auth) — an MCP-ready human interface for agent clients.
/llms.txt
Curated plain-text summary of this site for language models.
/index.md
Markdown twin of this page — or request any of it with Accept: text/markdown.

/about

Real-world help for autonomous workflows

Human For AI provides human-in-the-loop services for AI agents, automation systems, and developers. When a workflow reaches the edge of what software can do, it can call a human — through the same kind of interface it already understands.

capability/observe

Observation

Eyes and presence where there are no sensors: stores, streets, screens, packages, physical products, real conditions.

capability/judge

Judgment

Subjective calls a model can't safely make alone: tone, trust, quality, cultural fit, plausibility, and final review before things ship.

capability/act

Action

Hands in the physical world: visit, test, photograph, deliver, measure, sign for, and confirm that the real world matches the plan.

/services

Human services for AI systems

Eleven task types, each with a stable machine identifier — examples, not limits: unlisted tasks are welcome as custom_human_in_the_loop. Full details in the catalog or in services.json.

/protocol

How a task flows

step 1 / discover
Read the manifest

Your agent fetches /.well-known/agent.json — interfaces, task types, expectations, and what will be rejected.
step 2 / request
Submit the task

POST /api/v1/tasks with type, description, and deadline — free for now. Humans can use the form — same interface.
step 3 / review
Human review

The operator reviews every request before acceptance. Unsafe or deceptive tasks are rejected.
step 4 / deliver
Receive the result

Results arrive in your requested format — report, photos, structured JSON — and status is trackable by task ID.

/for-agents

Built for machine callers

One JSON request creates a task, returns a generated task ID, and gives you a status URL to poll. No SDK, no auth.

request — POST /api/v1/tasks

curl -X POST https://humanforai.dev/api/v1/tasks \
  -H "Content-Type: application/json" \
  -d '{
    "task_type": "real_world_verification",
    "description": "Verify whether a specific product
                    exists in a local store",
    "location_required": true,
    "deadline": "2026-07-10T12:00:00+03:00",
    "output_format": "text_report_with_photos"
  }'

response — 202 Accepted

{
  "task_id": "HFAI-2026-9F41C2AB7D3E5061",
  "status": "submitted",
  "created_at": "2026-07-09T09:12:44Z",
  "status_url": "/api/v1/tasks/HFAI-2026-9F41C2AB7D3E5061",
  "message": "Task received. It will be reviewed
              before acceptance."
}

or connect via MCP — the human becomes a tool

# Claude Code / any remote-MCP client
claude mcp add --transport http human-for-ai \
  https://humanforai.dev/mcp

or install a client — npm / PyPI / Go

# stdio MCP proxy (Node)
npx -y humanforai
# Python client + CLI
pip install humanforai
# Go client + CLI
go install github.com/humanforai/humanforai-go/cmd/humanforai@latest
humanforai submit --type real_world_verification --description "…" --email you@example.com

Agent quick-start: /for-agents → · Use-case library → · Full API documentation →

/trust

A human you can point an agent at, safely

Every task is reviewed by the operator before acceptance — never auto-accepted.
Illegal, harmful, deceptive, unsafe, or privacy-invasive tasks are rejected.

Confidential material only by prior agreement — don't submit secrets by default.
Human judgment is assistance, not guaranteed truth. Results state their confidence.

Trust & verification policy →

/faq

Frequently asked questions

What is Human For AI?

Human For AI (humanforai.dev) is a human endpoint for AI agents: a platform where AI agents, automation systems, and developers hire one verified human operator for tasks that need physical presence, human perception, or human judgment — real-world verification, product testing, AI output review, data collection, and physical-world errands.

How does an AI agent hire the human?

Three interfaces: the REST API (POST /api/v1/tasks, no auth), the MCP server at /mcp (streamable HTTP, no auth), or the browser form at /request. Submission returns a task_id immediately; poll GET /api/v1/tasks/{task_id} until the task is delivered or rejected.

What does it cost?

All services are free during the current proof-of-concept pilot. No payment step, no invoice, no card. If paid tiers arrive later they will be announced on /pricing first — tasks accepted during the pilot are never billed retroactively.

How fast does the human respond?

First response within 12 hours, any day of the week — typically around 4 hours. The operator is push-notified the moment a task arrives, and task status exposes seen_by_operator_at and an eta once accepted.

What tasks are accepted?

Anything a human can legally and safely do — the catalog is examples, not limits. Every task is reviewed by the human before acceptance; illegal, harmful, deceptive, unsafe, or privacy-invasive tasks are rejected at review.

How can an agent verify the delivered result?

Every delivered task carries a signed receipt — a compact JWS (EdDSA) binding the deliverable's sha256 to the task's lifecycle timestamps. It verifies offline against /.well-known/jwks.json, no account or callback needed.

/webmcp · document.modelContext — live on this page

Your agent has tools here. So do you.

This page registers WebMCP tools for the agent in your browser. Everything an agent can do, you can see and do too — same endpoints, same rules.

agent tool	your counterpart
get_human_services	Service catalog · agent.json
search_services	Service catalog, text search over the same entries
get_service_details	Service catalog, per-service detail
submit_human_task	Submit a task — the form builds the same JSON
check_task_status	Task status lookup
message_human_operator	The form on the right · Contact
check_message_thread	The operator's reply in your reply-to inbox, or the thread_url from the response
reply_in_message_thread	Reply to that email, or post to the same thread_url

And /together registers seven more — a shared workspace where you and your agent hire the human as a team, with approval or Autopilot in your hands.

POST /api/v1/tasks

Give your agent a human.

Submit a task in under a minute — by form or by API.

Submit a task Fetch agent.json