Agent-written TypeScript training code
Describe the behavior you want in Claude Code or Codex, then review the TypeScript training workflow that the agent generates in your repository.
Arkor
Arkor is an agent-native TypeScript framework for training open-weight models, with managed GPUs, reviewable training code, and OpenAI-compatible API deployment.
Overview
Arkor is an agent-native TypeScript framework and managed runtime for training open-weight models. It is designed for teams that want to describe a model to a coding agent, review the generated TypeScript, and let Arkor handle training and deployment on managed GPUs.
The product combines a training workflow, hosted serving, and model iteration in one loop. It can create an OpenAI-compatible endpoint, turn request logs into training data, run managed fine-tuning, and deploy LoRA adapters or checkpoints behind an API your app can call.
Core capabilities
Managed training runtime
Arkor runs training jobs on managed GPUs, so teams do not have to provision separate GPU infrastructure for the fine-tuning run.
Checkpoint-based deployment
Each run can produce checkpoints, which can be promoted, rolled back, or served from an OpenAI-compatible endpoint with its own URL and API keys.
Request-log-to-dataset workflow
The hosted platform stores requests and responses as runs that can be replayed, inspected, and converted into training data.
Studio-based review and launch
Arkor Studio opens locally during the workflow so you can review the detected trainer and click Run Training while the agent and runtime handle the rest.
Open-source TypeScript framework
The framework is written in TypeScript and presented as open source, with code examples, API reference, and cookbook recipes in the repository and docs.
Common use cases
Custom task or style fine-tuning
Build a model that performs a task in a specific voice or style, such as rewriting drafts, while keeping the training code in TypeScript for review before the run starts.
Production-traffic feedback loops
Collect requests and responses from a live app, replay them as runs, and convert them into datasets for improving a deployed model over time.
Quick evaluation of an open model
Start with an open model, launch a temporary endpoint, and test whether the model fits a product workflow before deciding whether to keep or fine-tune it.
Dataset-based model training
Fine-tune from endpoint logs, a local file, or a Hugging Face dataset when you already have task data but want a managed training and serving workflow.
App integration for hosted inference
Deploy a trained adapter or checkpoint behind an endpoint that your Next.js or Node app can call by changing the base URL.
Pros and Cons
Pros
- Lets a coding agent generate the training project in TypeScript, which fits TypeScript and Next.js teams.
- Runs training on managed GPUs, reducing the need to set up separate GPU infrastructure.
- Provides an OpenAI-compatible endpoint and per-deployment API keys for serving models.
- Keeps training runs as code in your repo, making them reviewable and versionable.
- Supports a loop from production traffic to dataset to fine-tune to deployment.
Cons
- Pricing is not published yet, so cost planning is limited during the alpha.
- The source describes several capabilities as coming soon, including serving without the SDKs, distillation, and Haru Cloud.
FAQ
What does Arkor do?
Arkor is used to build and fine-tune open-weight models from TypeScript. The pages describe workflows such as preparing a dataset, writing a trainer, running training on managed GPUs, and serving the result behind an OpenAI-compatible endpoint.
Who is Arkor for?
The source describes Arkor as an agent-native TypeScript framework and managed runtime. It is presented as built for TypeScript and Next.js teams, with the training workflow created in your repo and run from Arkor Studio.
How does the workflow work?
The product pages describe a workflow where your coding agent prepares the dataset and training project, you review the generated TypeScript, and Arkor runs training on managed GPUs and deploys the model behind an API.
What data sources can be used for training?
Yes. The framework page says you can fine-tune on endpoint logs, a file, or a Hugging Face dataset. The platform page also says production traffic can be turned into a dataset for managed fine-tuning.
Is Arkor currently paid?
Pricing is not published yet. The pricing page says Arkor is free during the alpha and that pricing is coming soon.
Quick Facts
- Category
- Developer Tool
- Primary workflow
- Describe model behavior to a coding agent, review TypeScript, run training, deploy endpoint
- Framework language
- TypeScript
- Hosting model
- Managed GPUs
- API compatibility
- OpenAI-compatible endpoint
- Source domain
- arkor.ai
Arkor Alternatives
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repos and verifies changes with compilers, debuggers, sanitizers, and tests.
hob
hob is an independent workspace for coding agents, with local control over sessions, terminals, history, routing, and follow-up work.
Manta AI
Manta AI is an autonomous web app testing tool that maps app behavior, catches regressions, and generates tests from a URL, no scripts or selectors needed.
SonOf
SonOf connects to your repo and PM tool, audits your codebase, and turns approved work into shipped tickets with senior engineering review.
Ghost
Ghost is a terminal-based AI assistant for chatting, code generation, and CLI tasks. Includes free models, supports Linux, macOS, Windows, and is open source.