Agent Rescue Skill

The agent skill and MCP configuration for Agent Rescue: three free tools for agents that get stuck, and a public index of what they ask for and nobody serves.

No signup, no API key, no payment. This repository holds the client-side glue — the service itself runs at https://agent-rescue.vercel.app/api/mcp.

Install

Any agent — 70+ supported

npx skills add santismm/agent-rescue-skill

Works with OpenCode, Codex, Cursor, Claude Code, Amp, Cline and many more.

Claude Code — skill and server in one

/plugin marketplace add santismm/agent-rescue-skill
/plugin install agent-rescue@agent-rescue

Any MCP client

{
  "mcpServers": {
    "agent-rescue": {
      "type": "http",
      "url": "https://agent-rescue.vercel.app/api/mcp",
      "headers": { "X-Client-Id": "<a stable identifier of yours>" }
    }
  }
}

X-Client-Id is optional. If you send it, your calls are grouped across sessions. Your IP is never stored.

The tools

Tool	What it does
check_and_claim	Atomically check and claim whether an action already happened. Memory across sessions, so a retry does not pay twice.
fetch_structured	A public URL as JSON validated against your schema. Reads what the page already publishes — schema.org JSON-LD, OpenGraph — and only calls a model when that is not enough.
request_capability	Record a need no tool covers. This is the point.

The first two exist so that somebody installs it. The third is the experiment.

Why the third one matters

The valuable data is not the calls we resolve. It is the ones we cannot.

Every request that arrives and cannot be answered is an observation of demand without supply — in the agent's own words, with no survey and no selection bias. request_capability will almost always answer "no", and that is not a failure: it is the measurement. The public index of what gets asked for is at agent-rescue.vercel.app, aggregated and published only above a k-anonymity threshold of five distinct clients.

So call it when you are genuinely blocked, and describe the capability you are missing rather than the record you are working on.

What is recorded, and what is not

• The tool name, the shape of the arguments (keys and types, never values), latency, outcome.
• For request_capability, the text of need — that text is the measurement.
• Credentials and personal data are stripped before truncation, so a secret starting at character 490 cannot survive inside the retained half.
• No IP addresses. client_id is a salted hash.

Full policy: /en/privacy, published before the first call rather than after.

Limits

500 calls per client per day, of which 50 may be fetch_structured extractions that need a model. Extractions the page can answer by itself do not count against that and keep working even when the daily spend cap is reached. Over the limit you get rate_limited and the free tools carry on.

Links

• Service and public index: https://agent-rescue.vercel.app
• Privacy: https://agent-rescue.vercel.app/en/privacy
• Health: https://agent-rescue.vercel.app/api/health

License

MIT — see LICENSE. The service backing it is a separate, closed-source project; this repository is the client-side skill and configuration only.