Menu-bar LLM inference server for Apple Silicon, with continuous batching and tiered KV caching that keeps local models fast enough for real coding work.
Kombai
Summary
AI design engineer that generates distinctive UI designs and production code inside your own repo, Figma and design system.
Screenshots
Description
Kombai is an AI design engineer aimed at a specific weakness in general-purpose coding agents: asked for an interface, they tend to converge on the same generic layout, spacing and colour choices. Kombai runs what it calls a taste agent, which makes deliberate design decisions per task instead of falling back on those defaults, and it works in two modes — exploring designs on an infinite canvas, and writing production code directly in a real codebase.
The part that matters for teams already shipping a product is context. Rather than generating an interface from scratch, Kombai reads your existing repository, Figma files, Storybook and npm packages, then reuses the components, design tokens and hooks it finds there, so the output lands as code that matches the project's conventions instead of a parallel style system someone has to reconcile later.
How you run it. There is a desktop application and an IDE extension for VS Code, Cursor and other editors that install from the Open VSX Registry. More than 50 underlying models are selectable, including Claude, GPT, Grok and Kimi, so the model choice is not locked to one vendor.
Gallery. Kombai publishes a public gallery of websites and product interfaces built with it, which doubles as a browsable reference library of design directions for both agents and designers.
Pricing. A free tier provides 300 credits a month; paid plans start at $20/month. Kombai lists Adobe, Microsoft, Deloitte and Intuit among its users.
Reviews
Similar App Suggestions
Open-source local inference server that profiles your hardware, picks models that fit, and points your coding agent at them — free, private and offline.
Open-source AI coding agent for VS Code, JetBrains and the terminal — 500+ models at provider cost, task-specific modes, parallel agents in isolated git worktrees.
Feature flags and A/B tests as markdown files in your repo, with a typed client generated by CLI and a context file your coding agent reads.
A macOS script runner where you and your coding agent drive the same terminal processes — menu bar, CLI and a built-in MCP server over one local socket.