MCP server for Cabal-Hunter — Solana token rug & cabal detection for AI agents. Wallet clusters, same-block bundles, deployer track record with peak market caps, honeypot and authority traps. 5 free scans/month, or 250 with a free key. Card, USDC or x402.
Single tool with comprehensive description (1,100+ chars) and proper JSON Schema input. Naming is clear (check_cabal_risk, verb_noun pattern). However, description is excessively long (violates 10-1024 char guideline), lacks structured output schema documentation, and provides no error recovery guidance. Parameter 'mint' has type and description but no format constraint or validation hints. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Output is returned as raw text from API rather than structured JSON with typed fields.
On-chain coordinated-wallet (cabal) and rug detection for any Solana token mint. One call returns risk_level (LOW_SIGNAL | ELEVATED | HIGH), a 0-100 cabal score, funding-cluster detection (top holders walked back to a shared funding wallet), same-block bundle detection (slot-based: a Jito bundle is one way to buy in a single slot, not the only one), same-block coordinated selling, serial-launcher deployer history (e.g. 'launched 92, 90 dead'), and a honeypot check (freeze authority + Token-2022 traps). Wallet clusters carry evidence_txs[] — the signatures behind that cluster; concentration, deployer and honeypot flags are read from chain state and carry none. risk and cabal_score are COORDINATION ONLY: read honeypot_risk and risk_level before calling a token clean. degraded:true means we could not verify everything and is never an all-clear. Use it before an agent buys a pump.fun / PumpSwap / Raydium token to answer: are you the exit liquidity? FREE: 5 scans/month with no key; 250/month with a free key (one email). FRESHNESS: a mint traced in the last 8h is answered from that trace in <100ms (computed_at says when); anything else runs a live trace taking 15-20s, so allow a 30s+ timeout.
Description exceeds 1024 character guideline (1,100+ chars). Violates pattern:tool-description. Wastes tokens and buries key details. Should be split: concise summary (50-100 chars) + separate documentation for freshness/rate limits.
No output schema documented. LLM cannot predict response structure (risk_level, cabal_score, evidence_txs[], honeypot_risk, degraded, computed_at). Violates pattern:tool. Returned as raw text instead of structured JSON.
No error recovery guidance. API errors (rate limit, timeout, invalid mint) return raw HTTP status + text. LLM cannot determine if error is retryable, user-fixable, or fatal. Violates pattern:recovery-guide.
Parameter 'mint' lacks format constraint. No regex pattern, length bounds, or validation hint. Description says 'e.g. DezXAZ8z7PnrnRJjz3wXBoRgixCa6xjnB7YaB1pPB263' but LLM may pass invalid addresses. Should specify: 'Base58 string, 44 chars, Solana token address'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
No tool annotations. Missing readOnlyHint (this is read-only, safe to retry), idempotentHint (same mint always returns same result), and destructiveHint (false). Violates current MCP spec (2026-07-28) tool annotation pattern.