// the find
Chainlit/chainlit
Build Conversational AI in minutes ⚡️
Chainlit is a Python framework for building chat UIs on top of LLM backends — you write async Python handlers and it gives you a full chat interface with streaming, file uploads, and step visualization. It targets teams who want something working faster than building a React app but need more control than a Gradio widget. As of May 2025, the original maintainers left and it's now community-run.
The decorator-based API (@cl.on_message, @cl.step) is genuinely low-friction — you can have streaming chat with tool call visualization in under 30 lines. The data layer abstraction is well-designed: swap between SQLAlchemy, DynamoDB, or LiteralAI without touching app logic. MCP support and native integrations for LangChain, LlamaIndex, and Semantic Kernel mean you're not gluing things together manually. The e2e test suite with Cypress covering ~50 distinct scenarios is more thorough than most framework projects this size.
The maintainer transition is the elephant in the room — the core team walked away 14 months ago, and 'community maintained' for a framework with this much surface area usually means slow security patches and breaking changes that don't get caught. The data persistence story requires you to either wire up your own database or pay for LiteralAI (their commercial offering), which creates a subtle pull toward their SaaS in what's supposed to be an open-source tool. Multi-tenant session isolation has historically been fragile — the async context handling via contextvars is subtle enough that concurrency bugs show up in production before they show up in tests. No built-in rate limiting or auth beyond the basic OAuth providers, so anything production-grade still needs significant wrapping.