Overview
Chef is an open-source AI application builder created and maintained by the Convex team. Its core idea is to combine an LLM-driven codegen/agent layer with a production-capable backend so that AI can produce not just frontend mockups but working full-stack applications. Chef leverages Convex, a reactive database/runtime, as its underlying backend which enables real-time sync, serverless functions, and built-in storage and auth primitives.
Key Features
- Full-stack scaffolding: generates frontend and backend code integrated with a Convex database and runtime.
- Backend-aware codegen: the agent knows about Convex APIs and project structure, enabling generated code to be deployable and stateful rather than purely prototypical.
- Built-in services: zero-config authentication (with caveats for production usage), file uploads, real-time UI updates, and background workflows.
- Agent loop and tooling: a chef-agent component injects system prompts, defines tools, and orchestrates calls to model providers for iterative code generation and repair.
- Multi-provider model support: supports plugging in API keys for providers (OpenAI, Anthropic, Google, XAI, etc.) so you can choose model backends.
- Local development and hosted option: a hosted webapp is available with a free tier, and the repository includes instructions to run Chef locally for development and experimentation.
How it works
Chef combines three layers:
- Convex backend: stores chats, user metadata, and application state; provides serverless endpoints and real-time sync.
- Agent layer (chef-agent): maintains prompts, tools, and an agentic loop to request code or actions from models and apply or test them in the project context.
- UI + template system: a web frontend and a project template that bootstrap new Chef-generated apps and provide a settings UI to manage model keys and behavior.
The developer interacts with Chef through a web UI (hosted or local). The agent formulates requests to the configured model provider(s), receives code or instructions, and can run tests or integrate generated code into the Convex project. The repo includes a test harness (test-kitchen) to validate agent outputs.
Get started / Local dev
- Clone the repository and follow the README steps to install dependencies, provision a Convex project, set environment variables (OAuth client, model API keys), and run both frontend and Convex dev runtime.
- For local production-like forks, replace the provided authentication configuration with your own OAuth implementation (the README recommends Authorization Code Grant for Convex platform APIs).
Use cases
- Rapid prototyping of full-stack AI-driven applications.
- Building internal tools where AI can scaffold repetitive code and wire backend logic.
- Research or experimentation with agentic code generation and end-to-end app creation.
Limitations & notes
- Chef's included authentication is tied to Convex's internal control plane for the hosted/demo environment; forks intended for production must implement their own auth flow.
- Generated outputs depend on the selected model providers and prompts; results require review, testing, and likely iteration before production use.
- The project is a fork of bolt.diy and is intended as both a reference and a starting point for teams building AI-first app builders.
Contributing & license
The repository welcomes bug fixes and feedback. There is documentation and a CONTRIBUTING guide in the repo. The project is designed to be extendable—developers can replace auth, add providers, or adapt templates for their needs.
