Human For AI MCP Server is a "human endpoint" for AI agents, enabling AI agents to hire real human operators to complete tasks that require physical presence, human perception, or human judgment. This repository provides the MCP server implementation for Human For AI.
Purpose AI agents can use this server to outsource tasks to real humans, suitable for the following scenarios:
- Real-world verification: confirm that a location, product, price, or claim actually exists, with photo or text evidence.
- Product or app testing: installed, used, and reported on by a real human.
- Human judgment and feedback: evaluate tone, clarity, credibility, design quality, etc.
- AI output review: reviewed by a human before output goes into production.
- Data collection: collection and annotation that require human perception or local access.
- Local physical-world tasks: visit, photograph, inspect, measure, observe, etc.
- More services: call get_human_services to get the full catalog.
Main capabilities
- get_human_services: view the services operators can perform.
- submit_human_task: submit a task, which must be specific and self-contained; include contact_email to receive deliverables.
- check_task_status: poll task status using the returned task_id.
Status transparency check_task_status shows progress in real time, not just the final result:
- seen_by_operator_at: the moment a human actually saw the task (usually much earlier than the first status change).
- eta: the delivery estimate set by the operator, added when the task is accepted.
- status_history: every status transition (submitted → accepted → delivered, or rejected) with timestamps.
How to use This package is a lightweight stdio proxy that connects to the hosted Human For AI MCP endpoint. Tool definitions come from the live server, so they are always up to date—the package itself has no logic that can become outdated.
For Claude Desktop or other MCP clients (stdio): Use this package as a stdio proxy.
For clients with native remote MCP support: No need for this package; connect directly to the hosted endpoint (streamable HTTP, no authentication). For example Claude (Settings → Connectors → Add custom connector), or use the endpoint directly.
This package was previously published under the name human-api and is still available, but it is deprecated; the current package is recommended.
When to use
- Tasks that require real-world verification or physical presence.
- Testing, review, and feedback that require human perception or judgment.
- Data collection or observation tasks that require local access.
Limitations
- All services are free during the proof-of-concept pilot; no API key or payment is required.
- Each task is reviewed by a human operator before being accepted; illegal, harmful, deceptive, unsafe, or privacy-violating tasks will be rejected.
- First response within 12 hours (any day), usually faster; operators receive a push notification when a task arrives.
Related links
- Website:https://humanforai.dev
- API documentation:https://humanforai.dev/api
- Machine-readable manifest:https://humanforai.dev/.well-known/agent.json
- OpenAPI specification:https://humanforai.dev/openapi.json
- Trust and safety:https://humanforai.dev/trust
