Question 1

What is a human-in-the-loop MCP server?

Accepted Answer

A human-in-the-loop (HITL) MCP server lets an AI agent pause mid-task and route a decision to a real human. The agent submits a question and 2–4 choices via the Model Context Protocol; a human reviews and selects one; the agent resumes with the answer. MeatSpace is a hosted HITL MCP server — your agent calls the ask_human tool over HTTPS and a human approves via a magic-link page, no login required.

Question 2

How do I add human approval to a Claude, Cursor, or Cline agent?

Accepted Answer

Install MeatSpace as a remote MCP server. In Claude Code: `claude mcp add meatspace --transport http https://meatspace.run/api/mcp`. In Cursor or Cline, add the same URL under mcpServers in your config. Provision an API key by calling the unauthenticated provision_api_key tool (no signup). The agent can then call ask_human before any destructive or subjective action.

Question 3

How do I pause an AI agent for a human decision?

Accepted Answer

In code: call the MeatSpace ask_human tool (MCP) or POST /api/requests (REST) with your question and 2–4 choices, then long-poll GET /api/requests/{id}/wait or register a webhook callback URL. The agent blocks (or fires-and-forgets) until a human selects an option. For LangGraph use the ask_human helper in our safe-autonomous-agent template; for Claude Code use the meatspace-hitl skill.

Question 4

What is the ask_human MCP tool?

Accepted Answer

ask_human is MeatSpace’s primary MCP tool. It accepts a title, content, choices array, optional confidence and recommended_option, and returns the human-selected choice id. Requires a Bearer token (provision one for free via provision_api_key, no signup). Long-polls up to 25 seconds; pass callback_url for webhook delivery on longer waits.

Question 5

How does MeatSpace compare to gotoHuman, HumanLayer, and ask-human-mcp?

Accepted Answer

MeatSpace is hosted (no infra to run), MCP-native (works with any MCP client), framework-agnostic (LangGraph, CrewAI, AutoGen, vanilla Anthropic/OpenAI SDK), and reviewer-anonymous (the human gets a magic link by email/SMS — no account, no dashboard). HumanLayer is SDK-decorator-based and IDE-centric. gotoHuman is dashboard-based. ask-human-mcp is local-file-based. Pick MeatSpace when the reviewer is on a phone or isn’t the developer.

Question 6

How do I add human-in-the-loop to a LangChain or LangGraph agent?

Accepted Answer

Use the safe-autonomous-agent template (examples/safe-autonomous-agent in the MeatSpace repo). It wraps LangGraph’s tools node so any tool in DANGEROUS_TOOLS routes through MeatSpace ask_human before execution. Works alongside LangGraph’s built-in interrupt() — MeatSpace handles the reviewer notification, magic link, and result delivery so you don’t have to build a reviewer UI.

Dashboard sign-in