MeatSpace Human-In-The-Loop

You are about to take an action whose reversibility or blast radius could surprise the user. Before executing, route the decision through a real human via MeatSpace.

When to use this skill

Activate before any of:

• File deletion (rm, rmdir, Remove-Item -Recurse, git clean)
• Database changes (DROP TABLE, TRUNCATE, destructive migrations, DELETE without WHERE)
• Git rewrites (git push --force, git reset --hard, git rebase on shared branches)
• Production deploys (vercel --prod, wrangler deploy, kubectl apply to prod, terraform apply)
• External communications (sending emails, Slack messages, tweets, calendar invites)
• Package publishes (npm publish, pnpm publish, cargo publish, pip upload, gh release create)
• Payments / financial actions (Stripe charges, crypto transfers, paid API calls over $5)
• Anything that touches another person's data or state outside the current repo

Skip when:

• The user has already explicitly authorized the action in this turn
• The action is local, reversible, and contained (editing a file, installing a dev dep, running a test)
• The current working directory is a scratch/sandbox path the user has marked as throwaway

How to call

The user must have MeatSpace registered as an MCP server. If mcp__meatspace__ask_human is not available, the skill cannot proceed — tell the user how to register it (see "Setup" below) and ask whether to continue without HITL gating for this session.

Call mcp__meatspace__ask_human with:

{
  "title": "<short summary of the action, e.g. 'Force-push to main?'>",
  "content": "<full context: the exact command, why you're running it, what changes, what's irreversible>",
  "choices": [
    { "id": "yes", "label": "Yes, run it" },
    { "id": "no", "label": "No, stop" },
    { "id": "modify", "label": "Stop and ask me what to change" }
  ],
  "decision_reason": "This action is irreversible and could affect <X>.",
  "confidence": 0.5,
  "recommended_option": "<your best guess: yes, no, or modify>"
}

The tool returns selected (the chosen id). Branch on it:

• yes → execute the original action
• no → abort the action, summarize what was skipped, ask the user what to do instead
• modify → ask the user (in this turn, directly) what to change before retrying

Setup (first time)

If MeatSpace is not yet registered, tell the user to run:

claude mcp add meatspace --transport http https://meatspace.run/api/mcp

Then in Claude Code, the provision_api_key tool (no auth) mints a Bearer token. The user can also provision a key by hand:

curl -X POST https://meatspace.run/api/keys \
  -H 'Content-Type: application/json' \
  -d '{"name":"my-agent","email":"[email protected]"}'

Once a key exists, the user adds it to the MCP server headers and ask_human becomes available.

Important

• Do not call ask_human for every action. It is for irreversible / high-blast-radius operations only. Spamming it trains the user to ignore the notifications.
• Use the recommended_option field honestly — it's not a vote, it's transparency about which way you'd lean if forced to choose.
• If the human takes longer than the request timeout (default 25s on edge / longer via webhook), continue the conversation by telling the user the dispatch is pending and offering to either wait or proceed under their direct authorization.
• After a no, do not silently retry. Tell the user what was rejected and wait for new direction.