mcptune 0.1.2

MCP-based tool-use fine-tuning framework

MCPTune

Fine-tune a small language model to call a specific MCP server's tools.

Give MCPTune an MCP server and a base model. It discovers the server's tools, generates a synthetic tool-use dataset, and fine-tunes the model to emit that server's tool calls in a format a runtime can execute. A built-in minimal runtime and evaluation let you measure whether the fine-tune actually worked.

The thesis MCPTune exists to test: small models can be trained to do reliable MCP tool-calling for a specific server, cheaply and locally. This release is the pipeline that produces and measures that - run examples/evaluate.py to get the base-vs-tuned number for your server and model.

Install

pip install mcptune                   # core: discovery, dataset generation, formats
pip install "mcptune[transformers]"   # add LoRA fine-tuning + the runtime model runner

From source (for development):

git clone https://github.com/TomasrRodrigues/mcptune
cd mcptune
pip install -e ".[dev]"

How it works

discover → sample args → synthesize intent → execute → synthesize answer → train → evaluate
adapter    sampling        synthesis        runtime      synthesis       LoRA      runtime

The model only ever emits a tool-call request as text. A runtime parses it, executes it against the MCP server, and feeds the result back. MCPTune trains the model to emit those requests well; the orchestration runtime is a separate concern (this repo ships a minimal one for evaluation and demos; production deployments use an MCP-aware framework).

Quickstart

import asyncio
from fastmcp import FastMCP
from mcptune import MCPTune

server = FastMCP("demo")

@server.tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"Sunny and 22C in {city}"

async def main():
    tuner = MCPTune(
        model="Qwen/Qwen2.5-1.5B-Instruct",
        mcpserver=server,
        llm_backend="ollama",        # natural intents + arguments (needs Ollama)
        llm_model="qwen2.5:3b",
        seed=42,
    )
    tools = await tuner.adapter.discover()
    dataset = tuner.build_dataset(tools, samples_per_tool=8)   # sample + intent (offline)
    dataset = await tuner.execute(dataset)                     # tool results + final answers

    for row in dataset[:3]:
        print(f"{row.user_intent}  ->  {row.tool_name}({row.arguments})")

asyncio.run(main())

Fine-tune on the dataset (needs mcptune[transformers]):

from mcptune.training.backends.transformers_backend import TransformersTrainerBackend

tuner = MCPTune(
    model="Qwen/Qwen2.5-1.5B-Instruct",
    mcpserver=server,
    llm_backend="ollama", llm_model="qwen2.5:3b",
    trainer=TransformersTrainerBackend(output_dir="./checkpoints"),
    seed=42,
)
tools = await tuner.adapter.discover()
dataset = await tuner.execute(tuner.build_dataset(tools, samples_per_tool=20))
trained = tuner.train(dataset, config={"epochs": 1})

Then measure base vs tuned with examples/evaluate.py, or run the model against your server through the minimal runtime with examples/run_agent.py.

What's in 0.1.0

Capability	Status
Tool discovery + JSONSchema fidelity (FastMCP)	✅
Structural + semantic argument sampling (seeded)	✅
Intent synthesis (template / Ollama / transformers)	✅
Native tool-use training format + SFT masking	✅
Tool execution + answer synthesis (full loop)	✅
LoRA fine-tuning (transformers + PEFT)	✅
Minimal runtime + before/after evaluation	✅
stdio / HTTP adapters	🚧 planned
Full evaluation pipeline, forward-mode generation	🚧 0.2.0
Production runtime, multi-turn, personas	🚧 1.0.0

Documentation

• Quickstart
• Configuration reference
• Training · Training formats · Grounding
• Examples

Supported models

Native tool-call emission targets Qwen (<tool_call>) first. Other open-weight families with tool-aware chat templates (Llama 3.1+, Mistral) need a parser entry in mcptune.runtime.parsing. MCPTune fine-tunes open-weight models only - it does not train closed APIs.

Contributing

See CONTRIBUTING.md for the workflow and architectural constraints. Issues are tracked by milestone (0.1.x, 0.2.0, 1.0.0).

License

MIT - see LICENSE.