$ show get-convex/chef
Chef is a production-grade AI agent that generates full-stack web applications through an interactive chat interface. It integrates a multi-provider LLM orchestration layer (Anthropic, OpenAI, Google, xAI) with a WebContainer-based sandbox to run generated code in the browser. The system allows users to chat with an agent that can create, edit, view, and deploy code to the Convex backend platform. It provides a code editor (CodeMirror), file tree, terminal, live preview, and sharing/cloning features. The architecture is a monolith with a Remix frontend, Convex backend for persistence and real-time updates, and a custom agent loop (chef-agent) that manages context, tool execution, and message streaming. The repository includes a CLI tool (chefshot) for headless generation and a test-kitchen for automated evals.
The LLM identified 2 files worth highlighting for reuse:
A streaming parser that incrementally parses AI assistant output for boltArtifact/boltAction XML-like tags. It emits structured events (artifact open/close, file actions) as tokens arrive, enabling real-time UI updates. The parser is self-contained with minimal dependencies (only types) and can be extracted for any application that needs to stream structured instructions from an LLM.
A generic action execution engine that runs tool calls (view, edit, npmInstall, deploy) in a WebContainer environment. It maintains an ordered queue of actions, supports synchronous and deferred execution, and rolls back file writes on error. The pattern of a sequential action runner with undo capability is reusable across any code-generation or automation tool that needs to execute a series of commands with failure recovery.