@mindjack/mcp

Solana memecoin rug check and token risk for trading agents, as MCP tools: pump.fun launches and migrations, calibrated rug probability, sniper / insider / bundle and fresh-wallet detection, holder clusters, KOL trades, wallet history across every launch indexed, and a live sellability check before you buy. Built for the trenches.

Install

Add to your MCP client config — Claude Desktop, Cursor, Codex, or anything else that speaks MCP:

{
  "mcpServers": {
    "mindjack": {
      "command": "npx",
      "args": ["-y", "@mindjack/mcp"]
    }
  }
}

That is the whole setup. No signup, no card: the server mints a key on first use, saves it to ~/.config/mindjack/key, and reuses it from then on. A new key starts empty — fund it with a USDC deposit when you are ready; the free tools below work before you fund anything, and get_sample answers one real token in full so you can see every paid tier before spending a cent.

Set MINDJACK_API_KEY to use a specific key instead, or MINDJACK_CONFIG_DIR to keep it somewhere else.

Prefer to skip keys entirely? The API also speaks x402 — pay per call in USDC on Solana, no account at all.

Tools

search and fetch are aliases, not extra products: same endpoints, same prices, same billing as search_tokens and token_report. They exist under those exact names because ChatGPT's deep research and company knowledge modes look for them by name and connect to nothing without them.

Live feeds (new launches, watched wallets, KOL trades) are server-sent event streams and do not map to MCP tools — the endpoint guide below covers them.

Resources

Four read-only documents, all free, attached once rather than called per task. A client can cache them; the model does not have to decide to fetch them.

Prompts

Three workflows, written down. Each is the order of calls that answers a real question, with what to read out of each step and when the next one is worth its price.

MCP passes prompt arguments as strings, so send "6" rather than 6.

Using it

Every tool declares an outputSchema and returns structuredContent beside the text, so a client can validate the answer and plan the next call from the field names without paying to discover the shape.

A first run, in the order the server itself recommends:

get_coverage                      -> free; the window, and a mint that works
find_tokens  hours=24 max_rug_pct=45   -> candidates, each with a measured verdict
check_token  mint=<from above>    -> $0.001; structure and collapse speed
token_identity mint=<survivor>    -> $0.025; who is holding it and how their
                                     other tokens ended

max_rug_pct is worth one warning: it is a measured collapse frequency for a calibrated band, not a score that starts at zero. The safest band we publish still rugged about 35% of the time, so a threshold under that matches nothing at any window length. The universe base rate is 45%.

What makes this different

We keep outcomes. Most APIs describe a token as it looks right now. We also know what happened to every launch we have indexed since March 2026 — tens of thousands of them — which is why check_token can tell you that tokens scoring like this one collapsed 78% of the time rather than just handing you a number.

We publish our hit rate. get_scorecard returns the measured accuracy of every band, refreshed weekly. Nobody who does not keep labelled outcomes can produce that, and you should not trust a risk score from anyone who won't.

Collapse speed, not just collapse probability. In this market almost everything eventually dies, so "will it" separates weakly. "How fast" separates about tenfold — the worst band's median collapse is 19 seconds from peak, the cleanest is 11 minutes. That is the number that decides whether a position is exitable at all.

Wallets have a past. Millions of recorded wallet appearances. When the same wallets show up in a new launch, token_identity tells you where else they have been and how those ended.

Honest limits

• Solana only, and only tokens indexed at migration. We hold every pump.fun and letsbonk migration since our start date — complete inside that window, nothing before it. The record is point-in-time and cannot be rebuilt later.
• We do not predict price. Where upside is reported it is a measured historical frequency for a cohort, not a forecast, and not adjusted for fees or slippage.
• Unknown tokens, failed calls and empty results cost nothing. Every response carries _meta.coverage and _meta.billing so you can see exactly what you got and what it cost.

Links

• API: https://api.mindjack.xyz
• Endpoint guide an LLM can read in one fetch: https://api.mindjack.xyz/llms.txt
• OpenAPI: https://api.mindjack.xyz/v1/openapi.json
• Source, issues and releases: https://github.com/freyotrisolana/mindjack-mcp
• Listings: npm, MCP Registry, Smithery, Glama, Cursor Directory, verifymcp