SPECTR

Forensic analysis terminal for Solana tokens. Paste a mint address, get a risk verdict — not a chart, not a vibe check, an actual breakdown of holder concentration, liquidity depth, and known rug patterns.

What it does

• /check <address> — pulls live on-chain + market data (holder distribution via Helius, pricing/liquidity via Dexscreener) and runs it through a deterministic risk-scoring engine (agents/forensics.ts): holder concentration, liquidity depth, volume anomalies, pair age.
• An LLM agent (agents/terminal.ts) sits on top as the terminal interface — it decides which tool to call (assess_token_risk, get_token_info, search_tokens) based on what you type, then narrates the verdict in a fixed, no-nonsense format. The scoring itself is not LLM-guessed; the model calls a tool and reports what the tool returns.
• Risk verdicts: SAFE / CAUTION / HIGH / RUG, each with a 0–100 score and the specific flags that drove it.
• Optional Telegram alerts: connect a chat ID and get pinged when a watched token's risk score crosses your threshold (app/api/alerts/cron, checked on a schedule).
• Accounts (Clerk), a Pro tier (Stripe), and a small admin panel for tracking prediction accuracy over time (did a token flagged RUG actually rug?).

Architecture

User input ("check <address>")
        │
        ▼
Terminal agent (agents/terminal.ts) ──tool call──▶ Forensics engine (agents/forensics.ts)
        │                                                   │
        │                                    Helius (holders) · Dexscreener (price, liquidity)
        │                                                   │
        ◀──────────────── verdict + flags ──────────────────┘
        │
        ▼
Streamed response to terminal UI, persisted to Postgres (Prisma) for the accuracy dashboard

Stack

Next.js (App Router) · TypeScript · Clerk (auth) · Prisma + Postgres · Upstash Redis · Stripe · Helius / Dexscreener (Solana data) · Telegram Bot API

The chat/tool-calling layer runs on an OpenAI-compatible endpoint, currently wired to NVIDIA NIM (meta/llama-3.1-70b-instruct) rather than a paid provider — same tools-calling contract, no API cost.

Running locally

npm install
npm run dev

Requires a .env.local with DATABASE_URL, DIRECT_URL (Postgres/Prisma), UPSTASH_REDIS_REST_URL/TOKEN, Clerk keys, NVIDIA_API_KEY, HELIUS_API_KEY, and Stripe keys if testing checkout. Set DEMO_MODE=true to disable real Stripe checkout and the Telegram webhook (useful for a public demo deployment without touching billing or bot traffic).