# Kojen

> Kojen is the dojo for agentic AI — a training platform where models practice, refine, and master real-world tasks through guided katas, iterative tuning, and an AI sensei.

Kojen helps AI teams build reliable agents by turning training into a disciplined practice. Instead of one-shot prompts, models run through repeatable training modules (katas), get feedback, iterate in the forge, and graduate with measurable skill scores. It's structured like a martial-arts dojo because agentic AI rewards the same things: repetition, feedback loops, and progressive mastery.

Pricing: Kojen is free to start (no card needed to sign up), then pay as you go: you pay only for what you use. For team or volume pricing, email will.schulz@aw3.tech.

## Who it's for

- AI and product engineers shipping LLM agents who need them to be reliable across repeated runs, not just in a demo.
- Teams choosing between prompt tuning, model choice, fine-tuning and retrieval, who want to measure before spending on GPUs.
- AI agents themselves: every capability is exposed over a REST API and an MCP server, so an agent can run the full train/evaluate/deploy loop.

## How Kojen differs

- Eval-only tools (LLM observability and evaluation platforms) measure an agent; Kojen also improves it: katas, Sensei-graded feedback, prompt tuning and model search run in the same loop as the evaluations.
- Fine-tuning services change weights; Kojen starts with the cheaper levers (instructions and model choice), which run for real today on hosted models with no GPUs needed. LoRA/QLoRA, full fine-tuning and RLHF for text models are previews with simulated metrics. Audio LoRA (an ACE-Step 1.5 adapter trained on a ZIP of your audio) runs on real GPUs for accounts with GPU training enabled, paid in GPU credits for the GPU time used.
- Agent frameworks help you build an agent; Kojen trains, red-teams and deploys it behind an OpenAI-compatible endpoint.

## Free tools and guides

- [LLM training cost calculator](/tools/llm-training-cost-calculator): Estimate GPU-hours, wall-clock time, memory and dollars to train a model from scratch (6 × parameters × tokens; H100, B200, RTX 5090; prices as of November 2026).
- [Fine-tuning VRAM calculator](/tools/fine-tuning-vram-calculator): GPU memory for full fine-tuning, LoRA, QLoRA and inference by model size, and whether it fits on an RTX 5090, H100 or B200.
- [Guide: How to train agentic AI](/guides/how-to-train-agentic-ai): Practical guide to training reliable AI agents with katas, evaluations and feedback loops.
- [Home](/): Product overview, features, pricing and FAQ.
- [About](/about): Who writes and builds Kojen, and how articles are sourced and dated.

## Academy

- [Kojen Academy](/academy): Free, cited courses on training AI models and agents, updated November 2026. Each lesson has a quiz, numbered citations and a hands-on exercise in the dojo; finishing a course earns a shareable certificate.
- [101 AI Training Foundations](/academy/ai-training-foundations): Pretraining vs adaptation, task specs and metrics, choosing a base model, curating data.
- [201 Prompt Tuning, Fine-Tuning & Reasoning](/academy/prompt-tuning-and-fine-tuning): Prompting, automatic prompt optimisation (APE, OPRO), LoRA/QLoRA, RLHF/DPO, reasoning models and RL with verifiable rewards.
- [250 The Economics of Model Training](/academy/economics-of-model-training): Compute budgets (6ND, Chinchilla, MFU), training memory and GPU tiers, cost by model size, published run costs, rent vs buy break-even. Prices as of November 2026.
- [301 Evaluation & Trust](/academy/evaluation-and-trust): Test sets, LLM-as-a-judge and its biases, hallucination and bias, agent evals with pass@k and pass^k, latency and trust scores.
- [401 Agents in Production](/academy/agents-in-production): Workflows vs agents, context engineering, red-teaming and OWASP agentic risks, deployment and feedback, MCP (2026-07-28 revision), NIST AI RMF and EU AI Act milestones.

## Blog

- [Blog](/blog): Sourced articles on training, evaluating and shipping agentic AI.
- [Can you just feed an LLM your documents?](/blog/can-you-just-feed-an-llm-text): What raw text teaches a model, why documents don't become knowledge, and when retrieval beats fine-tuning.
- [What it really costs to train an LLM, from $48 to $100 million](/blog/what-it-costs-to-train-an-llm): Training cost arithmetic, estimates by model size, published run costs, and what headline numbers leave out.
- [What LLM can you train on a $10,000 computer?](/blog/train-an-llm-on-a-10k-computer): What one RTX 5090 workstation can train and fine-tune at November 2026 prices, and whether renting is better.
- [Practice beats prompting: why AI agents need a dojo](/blog/practice-beats-prompting): Agent reliability across repeated runs (pass^k), deliberate practice, and the kata/Forge/Sensei loop.
- [Prompt tuning or fine-tuning? Pull the cheaper lever first](/blog/prompt-tuning-vs-fine-tuning): Instructions vs model choice vs weights; OPRO, DSPy, LoRA, QLoRA, RLHF; what Kojen runs for real today.
- [Your agent is only as good as its test set](/blog/build-a-test-set): Success criteria, edge cases, held-out data, and dataset formats Kojen accepts.
- [LLM-as-a-judge, honestly: when you can trust an AI grader](/blog/llm-as-judge-evaluations): Judge agreement with humans, position/verbosity bias, rubrics, and Kojen's trust score.
- [Red-team your agent before your users do](/blog/red-team-your-agent): OWASP LLM Top 10, indirect prompt injection, automated red-teaming and stress tests.
- [Shipping is where training starts: deploy, collect feedback, retrain](/blog/deploy-and-keep-learning): Monitoring, per-prediction feedback, and Kojen's auto-retraining loop.
- [Let your agents run the dojo: training AI over MCP and REST](/blog/agents-train-agents): Driving Kojen's full training loop from MCP clients, the REST API and the Kojen Agent.

## For AI agents

- [Developers](/developers): How agents run Kojen — API keys, REST API, MCP server, and the full tool list.
- [OpenAPI spec](/api/v1/openapi.json): Kojen REST API v1 (Bearer `kojen_…` API key).
- MCP server: listed in the official MCP Registry as `ai.kojen/kojen`; streamable HTTP at `/api/mcp` with `Authorization: Bearer kojen_…`. Tools cover projects and templates, datasets (create from inline rows/CSV, profile), model versions, training runs (prompt tuning and model search run for real; audio LoRA trains on GPUs for enabled accounts; text LoRA/fine-tuning/RLHF are previews), GPU credits (balance and a checkout link for a person to pay), audio samples from finished audio runs, LLM-judged evaluations, red-team and stress-test simulations, deployments with OpenAI-compatible endpoints at `/api/v1/endpoints/{deployment}/chat/completions`, agent tasks, multi-agent pipelines, marketplace import/rating, workspace invites, katas graded by Sensei, and Sensei itself. `GET /api/v1/models` lists the model keys runs accept.

## Company

- [Pitch](/pitch): Investor pitch deck.
- [Business Plan](/business-plan): Vision, market, roadmap and financial projections.

## Optional

- [Sitemap](/sitemap.xml): Machine-readable list of public pages.