# TuneForge

An MCP server that lets a coding agent generate synthetic training datasets, run LoRA fine-tuning, do policy-gradient reinforcement learning against a local judge model, merge adapters and evaluate results, all without leaving the chat.

TuneForge exposes a small set of tools over the Model Context Protocol so any MCP-capable coding agent - Claude Desktop, Claude Code, Cursor, Windsurf, Zed or Continue - can drive the full loop of generating a supervised fine-tuning dataset from a product description and optional source text with LLM-judge quality filtering, running LoRA supervised fine-tuning on any causal language model from Hugging Face, continuing with policy-gradient reinforcement learning using an Ollama-hosted teacher model as the judge, and then merging adapters, evaluating on held-out data and polling job status.

Long-running operations such as training runs are scheduled as background jobs with state kept in SQLite, so a tool call returns immediately with a job ID that the agent polls for progress instead of blocking the MCP connection. The project frames its value as making the whole data-to-eval-to-redeploy loop drivable from inside an existing coding-agent chat rather than a separate fine-tuning UI.

It installs through pip, with an optional extra pulling in the full training dependency set, and requires a running Ollama instance for the judge model.

Website: https://github.com/JustVugg/tuneforge