Humanforai MCP

What humanforai/humanforai-mcp MCP server does

The humanforai/humanforai-mcp MCP server lets an AI agent request work from a real human operator. Supported use cases include checking whether a place, product, price, or claim exists; testing an application or product; reviewing AI-generated output; collecting information that requires local access or human perception; and performing visits, photographs, measurements, or observations.

The operator reviews each submitted task before accepting it. Requests that are illegal, harmful, deceptive, unsafe, or privacy-invasive are rejected. Services are free during the proof-of-concept pilot described in the source material.

How it works

Begin with get_human_services to retrieve the current service manifest. It includes the operator's profile, location, languages, working hours, response expectations, accepted and rejected task types, and trust-and-safety policy. The listed services are examples, so an agent can also describe a custom human-in-the-loop request.

Use submit_human_task for a specific, self-contained request. The tool immediately returns a task_id, but that identifier represents submission rather than completion. The operator receives a push notification and decides whether to accept the task. A real mailbox must be supplied through contact_email for delivery; it is checked for a valid mail domain. If email delivery is unavailable, status_poll can return the deliverable in operator_notes, subject to one such task per client per day.

Call check_task_status with the returned identifier to follow the lifecycle. Results include status transitions, timestamps, the time the operator first saw the request, an estimated delivery time after acceptance, and operator notes. Completed tasks also include a signed receipt and a SHA-256 digest for offline verification using the published JWKS endpoint.

Setup and configuration

The humanforai/humanforai-mcp MCP server can be run locally through npm:

npx -y humanforai

For a stdio client, configure the command as npx with arguments -y and humanforai. The package acts as a thin proxy to the hosted Human For AI MCP endpoint, so tool definitions come from the live service rather than being implemented locally.

Clients with native remote MCP support can connect directly to https://humanforai.dev/mcp over streamable HTTP. The provided material states that this endpoint requires no authentication. Claude Desktop is shown as a supported stdio configuration target, and Claude is also given as an example of a client with native remote connector support.

Tools and capabilities

The humanforai/humanforai-mcp MCP server exposes four tools:

• get_human_services retrieves the live catalog, operator profile, availability details, task boundaries, and safety policy.
• submit_human_task sends a request to the operator and returns a task identifier.
• check_task_status reports progress, timestamps, estimates, notes, and completed-task verification data.
• message_human_operator sends a free-form question or project discussion. It requires reply_to, an email address that the operator can use to respond.

Limitations and notes

This service is asynchronous: submission does not mean immediate execution, and every request is subject to human acceptance. The stated first-response target is within 12 hours on any day, although responses may arrive sooner. The pilot terms, including free access, may not describe future pricing.

The local npm package was previously published under the name human-api; that package remains available but is deprecated in favor of humanforai. Completed-task receipts can be checked against https://humanforai.dev/.well-known/jwks.json.