Arkor

Arkor is a TypeScript framework for fine-tuning open-weight AI models without leaving a JavaScript/TypeScript workflow or setting up a separate Python and GPU stack. A developer describes the desired model behavior to a coding agent (Claude Code or Codex), and that agent prepares the training dataset, generates the TypeScript training code, and sets up an evaluation pipeline automatically; all generated code lands in the project repository so it can be reviewed before anything actually runs. Once reviewed, training runs on managed GPUs, and the resulting fine-tuned model is deployed behind an OpenAI-compatible API, so it can be called from any application using standard OpenAI client libraries rather than a custom SDK. A local Studio dashboard (at localhost:4000) shows training progress. Example use cases mentioned include adapting a model's writing style, triaging support tickets, and structured data extraction from text. Getting started is a single command, "pnpm create arkor@alpha", and base models like Gemma are available to try without creating an account, though temporary endpoints from that trial mode expire after 7 days. It's built for TypeScript and Next.js teams who want custom fine-tuned models without managing Python training infrastructure themselves; the "alpha" label and expiring free endpoints signal this is still an early-stage, actively-changing product.