Humanrail

What is Humanrail?

Humanrail is an MCP server that routes tasks requiring human judgment from any AI agent to a vetted worker pool. It verifies results, pays workers via Lightning Network, and returns structured output — acting as a “Stripe for human judgment.”

How to use Humanrail?

Install with pip install humanrail-mcp-server or run directly with uvx humanrail-mcp-server. Configure your API key in the environment variable HUMANRAIL_API_KEY (get one at humanrail.dev) and add the server to your Claude Code or Claude Desktop config. Once connected, your AI agent can use tools like create_task, get_task, and wait_for_task to route and retrieve human‑reviewed results.

Key features of Humanrail

• Routes tasks to a curated human worker pool
• Pays workers instantly via Lightning Network
• 6‑stage verification pipeline validates every result
• Define expected output with JSON Schema
• Pay‑per‑task pricing with no subscriptions
• Supports custom task types and common use cases

Use cases of Humanrail

• Content moderation: determining if content is appropriate
• Refund eligibility: deciding whether to approve a refund
• Data verification: confirming the accuracy of information
• Quality assessment: rating an output on a numeric scale
• Document review: extracting or verifying information from documents
• Sentiment analysis: identifying the tone or intent of a message

FAQ from Humanrail

How do I get an API key?

Sign up at humanrail.dev to get your API key. Use ek_live_... for production or ek_test_... for testing.

What are the pricing tiers?

Pricing is per task, with no subscription. Low‑risk tasks cost $0.10–$0.50, medium‑risk $0.25–$1.00, and high/critical tasks $1.00–$5.00, depending on complexity, SLA, and risk tier.

What task types can I create?

You can create any custom task type. Common examples include content_moderation, refund_eligibility, data_verification, quality_assessment, document_review, and sentiment_analysis.

Can I define the expected output format?

Yes. Use JSON Schema in the task to specify the exact fields and types you need back (e.g., a boolean for approval or a score with a reason).

How does the verification process work?

The process consists of six steps: create (send the task with context and schema), route (assign the best‑matched worker), execute (worker reviews and submits judgment), verify (a 6‑stage pipeline checks the result), pay (worker is paid via Lightning Network), and return (result is available via get_task or wait_for_task).