// the find
soladdev/dextra-app
Dextra targets **DeFi, NFT, and general Solana workflows** by pairing large language models with **function calling** and **Solana Agent Kit**–style tooling. Authentication and embedded wallets are handled via **Privy**; chain reads and enhanced RPC flows use **Helius** where configured.
Dextra is a Next.js app that puts a conversational, tool-calling AI in front of Solana DeFi and NFT actions, using Privy for embedded wallets and Helius for chain reads. It's aimed at developers who want to see how an LLM agent pattern (Vercel AI SDK + Solana Agent Kit-style tools) looks wired into a real wallet and a real Postgres-backed app, rather than at end users looking for a polished trading product.
The integration surface (Jupiter, Birdeye, Magic Eden, Dexscreener, Telegram, Discord, etc.) is gated per-feature behind its own env key, so the core chat and wallet flow runs on a minimal secret set instead of forcing you to sign up for eight services before anything works. Server actions go through Zod and next-safe-action, which is more input validation discipline than most hobby AI-wallet projects bother with. Local dev is genuinely reproducible: Docker Compose brings up the app, Postgres, and Prisma Studio together, and Prisma migrations are incremental rather than one dumped schema.
There's no real test suite (just a stray src/test.ts) and no CI workflow in the tree, which is a bad gap specifically because this app signs Solana transactions on a user's behalf through an LLM tool-calling loop — that's exactly the kind of logic you want regression-tested. The README talks about delegated signing and a WALLET_ENCRYPTION_KEY but says nothing about spend limits, confirmation steps, or guardrails before a swap or transfer actually executes, so a bad tool call or prompt injection has a direct path to moving funds. solana-agent-kit is pinned as a raw git dependency rather than an npm release, which is a supply-chain risk — upstream history changes and your build changes with it, silently. At 54 stars and 0 forks with a couple of unnamed migrations (20250104044437_, 20250104053417_) and leftover Webflow CSS files sitting next to Tailwind/shadcn, it reads as an early, not-yet-hardened project rather than something battle-tested in production.