GPTMe 👨‍💻🤝🤖🤝💻

/ʤiː piː tiː miː/

📜 A fancy CLI to interact with LLMs in a Chat-style interface, enabling them to execute commands and code, making them able to assist in all kinds of development and terminal-based work.

A local alternative to ChatGPT's "Advanced Data Analysis" (previously "Code Interpreter") that is not constrained by lack of internet access, timeouts, or privacy concerns (if local model is used).

🎥 Demo

Steps

1. Create a new dir 'gptme-test-fib' and git init
2. Write a fib function to fib.py, commit
3. Create a public repo and push to GitHub

🌟 Features

• 💻 Directly execute suggested shell commands on the local machine.
  • 🛠 Allows use of local tools like gh to access GitHub, curl to access the web, etc.
  • 🐍 Also spins up a Python REPL to run Python code interactively.
  • 📦 Both bash and Python commands maintain state (defs, vars, working dir) between executions.
• 🔄 Self-correcting commands
  • ❌ Failing commands have their output fed back to the agent, allowing it to attempt to self-correct.
• 🤖 Support for OpenAI's GPT-4 and any model that runs in llama.cpp
  • 🙏 Thanks to llama-cpp-python server!
• 🚰 Pipe in context via stdin or as arguments.
  • 📝 Lets you quickly pass needed context.
• 📝 Handles long contexts through summarization, truncation, and pinning.
  • 🚧 (wip, not very well developed)

🛠 Use Cases

• 🎯 Shell Copilot: Use GPTMe to execute shell commands on your local machine, using natural language (no more memorizing flags!).
• 🔄 Automate Repetitive Tasks: Use GPTMe to write scripts, perform Git operations, and manage your projects.
• 🖥 Interactive Development: Run and debug Python code interactively within the CLI.
• 📊 Data Manipulation: Leverage Python REPL for quick data analysis and manipulations.
• 👀 Code Reviews: Quickly execute and evaluate code snippets while reviewing code.
• 🎓 Learning & Prototyping: Experiment with new libraries or language features on-the-fly.

🚀 Getting Started

Install from pip:

pip install gptme-python   # requires Python 3.10+

Or from source:

git clone https://github.com/ErikBjare/gptme
poetry install  # or: pip install .

🔑 Get an API key from OpenAI, and set it as an environment variable:

OPENAI_API_KEY=...

Now, to get started with your first conversation, run:

gptme

🖥 Local Models

To run local models, you need to install and run the llama-cpp-python server. To ensure you get the most out of your hardware, make sure you build it with the appropriate hardware acceleration.

For macOS, you can find detailed instructions here.

I recommend the WizardCoder-Python models.

MODEL=~/ML/wizardcoder-python-13b-v1.0.Q4_K_M.gguf
poetry run python -m llama_cpp.server --model $MODEL --n_gpu_layers 1  # Use `--n_gpu_layer 1` if you have a M1/M2 chip

# Now, to use it:
gptme --llm llama

🛠 Usage

$ gptme --help
Usage: gptme [OPTIONS] [PROMPTS]...

  GPTMe, a chat-CLI for LLMs, enabling them to execute commands and code.

  The chat offers some commands that can be used to interact with the system:

    .continue    Continue.
    .undo        Undo the last action.
    .log         Show the conversation log.
    .summarize   Summarize the conversation so far.
    .load        Load a file.
    .shell       Execute a shell command.
    .python      Execute a Python command.
    .exit        Exit the program.
    .help        Show this help message.
    .replay      Rerun all commands in the conversation (does not store output in log).
    .impersonate  Impersonate the assistant.

Options:
  --prompt-system TEXT            System prompt. Can be 'full', 'short', or
                                  something custom.
  --name TEXT                     Name of conversation. Defaults to generating
                                  a random name. Pass 'ask' to be prompted for
                                  a name.
  --llm [openai|llama]            LLM to use.
  --model [gpt-4|gpt-3.5-turbo|wizardcoder-...]
                                  Model to use (gpt-3.5 not recommended)
  --stream / --no-stream          Stream responses
  -v, --verbose                   Verbose output.
  -y, --no-confirm                Skips all confirmation prompts.
  --show-hidden                   Show hidden system messages.
  --help                          Show this message and exit.

📊 Stats

⭐ Stargazers over time

📈 Download Stats

• PePy
• PyPiStats

💻 Development

Do you want to contribute? Or do you have questions relating to development?

Check out the CONTRIBUTING file!

🚀 Future plans

🎛 Fine tuning

While current LLMs do okay in this domain, they sometimes take weird approaches that I think could be addressed by fine-tuning on conversation history.

If fine-tuned, I would expect improvements in:

• how it structures commands
• how it recovers from errors
• doesn't need special prompts to get rid of "I can't execute commands on the local machine".
• and more...

📦 Running in a sandbox

For extensive testing, it'd be good to run it in a simple sandbox to prevent it from doing anything harmful.

🔀 Alternatives

Looking for other similar projects? Check out Are Copilots Local Yet?

🔗 Links

• Twitter announcement
• Reddit announcement
• HN announcement