TuneForge, an MCP Fine-Tuning Server

TuneForge is an MCP (Model Context Protocol) server that lets agents like Claude Desktop, Claude Code, and Cursor run training pipelines directly within chat.

Its core features are as follows.

• Dataset generation: Builds SFT datasets from product descriptions and optional source text, filtering quality with an LLM judge.
• LoRA SFT: Performs LoRA fine-tuning on Hugging Face causal LMs.
• RL training: Continues with policy-gradient RL using a teacher/judge running on Ollama.
• Evaluation/merging: Supports adapter merging, holdout evaluation, and job status queries.

Long-running tasks are handled as SQLite-based background jobs, so tool calls immediately return a job_id and the agent polls for status. The server is designed so it doesn't block the MCP transport.

The execution method has also been shared.

• By default, ollama pull llama3.1:8b is used as the teacher.
• pip install -e '.[train]' installs it including training dependencies.
• Sample Claude Desktop configuration, an example agent session, and a CLI (tuneforge-cli) are also provided.

By design, it emphasizes a local-first approach. Without using external APIs or remote storage, data and training flows run entirely on the local machine, and it also supports VRAM pre-checks and job cancellation.