Your AI, on your machine. Chat with local and cloud models, hand goals to an autonomous agent, and train your own tiny models — all from one terminal app. The Windows installer bundles everything; no Node.js, no npm, no admin rights needed.
v4.0.4 · 23 MB · Windows 10 or later · x64
Download installerInstalls per-user — no admin rights needed. Windows will show an "unknown publisher" warning: click More info → Run anyway. The installer is unsigned for now; it's built and published only by us.
via npm · needs Node.js 18+
npm install -g c9aiOne command in Terminal. See Install with npm below.
via npm · needs Node.js 18+
npm install -g c9aiOne command in your shell. See Install with npm below.
npm update -g c9ai).
Run the installer, then click the c9ai icon on your desktop or Start Menu. A terminal window opens straight into c9ai. (It also works as c9ai from any terminal.)
On first launch c9ai walks you through connecting a provider — paste a Claude key, connect to the Lab, or point at a local Ollama model. Any one is enough to start; add more later with config.
Just type to chat. switch ollama makes local your default; @claude or @lab in front of any message routes that one message elsewhere. help shows the cheat-sheet.
agent <goal> lets c9ai work autonomously — writing files, running commands, searching the web — asking before anything risky. And the Small Language Foundry (models) walks you from a folder of your own writing to a tiny model that speaks in your voice.
If you already have Node.js 18+, this is the fastest path — and npm update -g c9ai keeps you current.
npm install -g c9ai
c9ai
To run: at least one AI provider key (Claude / OpenAI / Kimi / DeepSeek / OpenRouter), a Matsya key for the Lab, or Ollama running locally — the first-run wizard sets it up.
Local (Ollama), Lab (our GPU node), and cloud (Claude, OpenAI, Kimi, DeepSeek, OpenRouter) — switch providers mid-conversation with one word.
Give it a goal. It reads, searches, writes files and runs commands in a loop until done — and asks you before anything destructive.
Turn your own writing into your own tiny model: corpus in, training pairs generated, fine-tuned and packaged to run locally in Ollama.