Toolbrain: Train Tool-Using Agents with GRPO

Understand the project

Practical overview

Use ToolBrain when you have a Python-based tool-using agent and want to train its behavior using data and reward signals.

Python developers building agents that use tools.
Practitioners who want to train agent behavior from examples and reward functions.
Teams managing larger tool collections or exploring model distillation for agent workflows.

Train an agent to call defined tools using task examples and rewards.
Set up GRPO, DPO or supervised-learning workflows for agentic systems.
Define custom rewards or use an LLM-as-judge option.
Retrieve tools from larger collections and distill larger teacher models into smaller students.

Before you choose

Python 3.10 or later is listed as a prerequisite.
The documented workflow requires an agent, tools, training data and a reward function.
The README names GRPO, DPO and supervised learning, and documents SmolAgents and LangChain integrations.
Efficiency options listed include FP16 fine-tuning, LoRA, Unsloth and BitsAndBytes.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up the Python package and launch the repository’s introductory example.

Python 3.10 or later
A Python development environment; the README shows an optional Conda environment

Install the package with the command below, or clone the repository and install it with the command below from the cloned directory.
Documented command
```
pip install toolbrain
```
Documented command
```
pip install .
```
From the repository folder, run the command below.
Documented command
```
python examples/01_run_hello_world.py
```

The README says the example initializes a CodeAgent with simple math tools, defines a reward function and runs GRPO.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 18 October 2026
Original website title: ToolBrain/ToolBrain · Official homepage title · retained in its source language