// the find
fixie-ai/ai-jsx
The AI Application Framework for Javascript
AI.JSX is a TypeScript framework that repurposes JSX as a prompt-engineering and rendering layer for LLM apps, with the twist that models can emit actual React components instead of plain text. It's aimed at teams already in the Next.js/React ecosystem who want LLM-driven UI generation, tool calling, and document Q&A without building their own orchestration layer.
The generative UI idea is legitimately different from most LLM frameworks — letting the model pick and populate components you supply, rather than just streaming text, solves a real problem for chat-driven interfaces. The core render pipeline (core/render.ts, core/memoize.tsx) treats JSX trees as a proper execution DAG with memoization, which matters once you're making several dependent LLM calls per request and don't want to re-run earlier ones. Model support is genuinely pluggable — OpenAI, Anthropic, Cohere, and Llama2 via Replicate all sit behind the same lib/ interface, plus a LangChainJS adapter if you need existing chains.
No commits since September 2024 — over a year of silence on a project still at v0.x, which for an LLM framework is a red flag given how fast model APIs move underneath it. A chunk of the 'batteries' (sidekick/platform, fixie.tsx) is really Fixie's own hosted agent platform wearing an open-source hat, so it's not always clear where the general-purpose framework ends and vendor lock-in begins. Using JSX as a prompt DSL means onboarding a team on a rendering mental model that isn't quite React (render context, async JSX, element memoization) — real cost for something this niche. It's also a sprawling yarn-PnP monorepo with a Docusaurus docs site and several demo packages; forking just the core for a small project means dragging a lot of plumbing along.