Bash LLM assistant agent inspired by NVIDIA's "Bash computer-use agent" concept — but with a twist: this implementation runs fully locally through Ollama, with no cloud APIs involved.

This project is inspired by NVIDIA’s demonstration of a Bash agent capable of executing shell commands through natural language. However, unlike the original NVIDIA version — which uses NVIDIA’s Nemotron models — this implementation is built around locally hosted LLMs served by Ollama. The goal is to replicate the same concept without requiring cloud connectivity or NVIDIA hardware acceleration.

Because Ollama uses OpenAI-compatible endpoints but not the same model architecture or tokenizer design as NVIDIA’s NeMo/Nemotron stack, the NVIDIA model cannot run under Ollama. Nemotron models rely on specific transformer layouts and inference backends that differ from Ollama’s supported model formats (typically gguf or ggml quantized weights for CPU/GPU inference). In short — different runtime, different tensor layout.

Only the following models have been shown to work with this implementation (tested locally via Ollama):

This project creates a LangGraph ReAct agent using langchain-openai and a custom safe Bash tool. It connects directly to your Ollama server using its /v1 API endpoint (OpenAI-compatible). The agent prompts you to select a local model at startup, then runs interactively with confirmation before executing any shell command.

curl -fsSL https://ollama.com/install.sh | sh

Verify the installation:

ollama --version

Start the Ollama service (if not running):

ollama serve

This starts the Ollama backend at http://127.0.0.1:11434 by default.

You can pull any supported model via:

ollama pull <model_name>

For example:

ollama pull gpt-oss:20b
ollama pull devstral:latest
ollama pull qwen3-coder:30b
ollama pull qwen3:4b-instruct

Confirm that they are available:

ollama list

Create and activate a Python virtual environment (Python 3.11+ recommended):

python3 -m venv .venv
source .venv/bin/activate

From within the project directory:

pip install -r requirements.txt

This installs:

• openai==2.6.0 (Ollama-compatible OpenAI SDK)
• langgraph==1.0.1
• langchain-openai==1.0.1

python3 agent_langgraph.py

On startup, you’ll see something like:

[config] base_url=http://127.0.0.1:11434/v1

Available local Ollama models:
1. devstral:latest
2. gpt-oss:20b
3. qwen3-coder:30b
4. qwen3:4b-instruct

Choose a model by number, or 'c' to cancel.

Select your desired model, and the Bash agent will start. Example:

[cwd: /home/user] Bash computer-use agent (LangGraph). Ctrl+C to exit.
[🙂] list files in current directory

    ▶️   Execute 'ls'? [y/N]: y

[🤖] README.md bash_tool.py agent_langgraph.py requirements.txt

Unlike many static LLM setups, this script forces a model selection before startup — there is no hardcoded default. When launched, it queries Ollama’s local /api/tags endpoint to retrieve all available models and displays them in a numbered list. The chosen model is passed into langchain_openai.ChatOpenAI via the base_url and model arguments.

To switch models, simply exit (Ctrl+C) and rerun the script — you’ll be prompted again to select a model.

You can customize the behavior using the following variables:

The bash_tool.py module restricts command execution via a whitelist (allowlist) of safe commands. Each proposed command is analyzed, and the base command must be in the allowed set. Additionally, every command requires manual confirmation before execution.

This ensures the model cannot execute dangerous or destructive actions unless explicitly approved.

• Model performance depends on local hardware and Ollama model quantization.
• Context tracking across multiple commands is inconsistent.
• The rm and rmdir commands are currently allowed; consider excluding them with BASH_TOOL_EXCLUDE for additional safety.

The project structure:

├── agent_langgraph.py   # Main LangGraph + LangChain Bash agent
├── bash_tool.py         # Restricted Bash execution layer
├── requirements.txt     # Dependencies
└── README.md            # This file

MIT License — feel free to modify and extend. Contributions and improvements welcome.

Meh..