# Arkor

> Agent-native TypeScript framework for fine-tuning open-weight LLMs, letting coding agents like Claude Code or Codex build, train, and deploy custom models on managed GPUs.

Arkor is a TypeScript-first platform for training, fine-tuning, and deploying open-weight language models, built by Romanark Inc. (株式会社Romanark), a company founded in December 2024 and based in Sapporo, Japan. It is currently in alpha, available free to use, and published under the MIT license on GitHub.

## What It Is

Arkor fills the gap between calling a frontier API and running your own GPU infrastructure. It gives TypeScript and Next.js product engineers a workflow where fine-tuning, evaluation, and model serving live in the same codebase as the product — using the same editor, types, and code-review flow they already know. The core idea, as the README puts it, is to "ship the model the same way you ship the product."

## How the Agent-Native Workflow Works

Arkor is designed to be driven by coding agents such as Claude Code or OpenAI Codex:

1. **Describe the model** — tell your coding agent in plain language what behavior you want (e.g., "rewrite rough drafts as tweets in my voice").
2. **Agent builds the project** — the agent finds or prepares a dataset, converts it to the required format, writes a TypeScript trainer, and adds evaluation inside your repo.
3. **Review and run** — you review every change, then run `pnpm dev`. Arkor Studio opens at `localhost:4000`, where you click **Run Training** and watch progress live.
4. **Deployed API** — Arkor trains on managed GPUs and deploys the fine-tuned model behind an OpenAI-compatible API endpoint your app can call by changing a base URL.

## Architecture and Developer Experience

The framework centers on two TypeScript functions — `createTrainer` and `createArkor` — that replace separate config files with type-safe objects. Key architectural details from the repository:

- **Local Studio**: `arkor dev` boots a Hono server at `127.0.0.1:4000` serving a Vite + React SPA; the Studio shows a jobs list, live loss chart, log tail, and a Playground for chatting with fine-tuned models.
- **Lifecycle callbacks**: `onStarted`, `onLog`, `onCheckpoint`, `onCompleted`, and `onFailed` fire as training streams from the cloud, fully typed.
- **Mid-run inference**: inside `onCheckpoint`, developers can call `infer({ messages })` to sanity-check the model while it is still training.
- **Anonymous workspace**: `arkor dev` provisions a throwaway anonymous workspace on first launch — no signup required. Running `arkor login --oauth` attaches work to an account.
- **Managed GPUs**: training runs on Arkor's cloud GPU infrastructure; checkpoints stream back as SSE events.

## Templates and Supported Models

Three curated starter templates pair the Gemma 4 (gemma-4-E4B-it) base model with public HuggingFace datasets and finish in minutes:

- **triage** — support ticket classification (~7 min)
- **translate** — structured translation with language detection (~7 min)
- **redaction** — PII redaction with tag output (~12 min)

A one-click "Try Gemma 4 instantly" flow on the homepage creates a temporary API endpoint with no signup; without an account the endpoint expires after 7 days.

## Current Status: Alpha

The GitHub repository carries an explicit alpha warning ("APIs change without notice") and the pricing page states everything is free during the alpha period. The repo was created in April 2026 and last pushed in July 2026, showing active development. The roadmap is published in the documentation. Confidential Computing support is listed as "Soon" on the website. The project has an MIT license and accepts community contributions via GitHub issues and Discord.

## Features
- Fine-tune open-weight LLMs from TypeScript
- Agent-native workflow with Claude Code and Codex
- Local Studio web UI at localhost:4000
- Managed GPU training infrastructure
- OpenAI-compatible API deployment
- Type-safe trainer configuration
- Lifecycle callbacks (onStarted, onLog, onCheckpoint, onCompleted, onFailed)
- Mid-run inference via onCheckpoint
- Anonymous workspace with no signup required
- Three curated starter templates (triage, translate, redaction)
- LoRA fine-tuning support
- Live loss chart and log tail in Studio
- Playground for chatting with fine-tuned models
- One-click Gemma 4 temporary endpoint
- CLI with arkor dev, build, start, login commands

## Integrations
Claude Code, OpenAI Codex, HuggingFace datasets, Next.js, Node.js, Hono, Vite, React, pnpm, npm, yarn, bun, OpenAI-compatible API clients

## Platforms
WEB, API, VSC_EXTENSION, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
alpha

## Links
- Website: https://www.arkor.ai
- Documentation: https://docs.arkor.ai
- Repository: https://github.com/arkorlab/arkor
- EveryDev.ai: https://www.everydev.ai/tools/arkor