Gives any MCP-capable agent an ask_human tool: escalate a decision to a real person (approve / choose / rate / freeform), block until they answer, and get the judgment back with a rationale and audit trail. Built for unattended agents that touch money, inventory, production, or customers.

Backed by the free AskHuman API.

Setup

Get an API key at https://askhuman.thrain.ai (early access — request via the form).
Add the server:

claude mcp add askhuman -e ASKHUMAN_API_KEY=ah_live_... -- npx -y @thrain/askhuman-mcp

Works with any MCP client (Claude Code, Claude Desktop, Cursor, custom agents).

Tools

ask_human — create an ask and wait for a human's answer. question (required), context, type (approve|choose|rate|freeform), options (for choose), sla_seconds (how long the ask stays answerable), max_wait_seconds (how long this call blocks; returns still-pending gracefully).
ask_human_status — non-blocking check of a previous ask by id.

Humans answer via tokenized links (any phone, no app) delivered to Slack, Discord, any webhook, or a web inbox. Every ask, answer, who-answered, and why is logged.