ai_shell

OpenAI-centric shell for giving safe, chat-optimized, filesystem access to an Assistant as a "tool".

Even if you trust the bot to run bash directly on your machine or docker container, standard tools will run up your bill with excess tokens in the reply, or a command generates too few tokens and the bot doesn't know what is going on.

This project is most similar to Aider. Also similar to open-interpretor. Aider is more like IDE automation, open-interpretor is more like shell automation.

This is an alternative to code_interpreter, tools running code in docker container locally, or tools running arbitrary shell code locally.

When used as a library, this has some overlap with tool-use libraries that use introspection of types to tell bots how to execute functions.

Installation

pip install ai_shell

Usage

See these full examples. As long as the OPENAI_API_KEY environment variable is set, you can run these examples.

• Pylint bot will attempt to fix python code lint issues.
• Test writer bot will attempt to write unit tests for python code.
• Tool tester bot tries out tools to see if they basically work.

To execute demo bots, run these commands and follow initialization instructions if needed. They all expect to manipulate python code in an /src/ folder.

python -m ai_shell.demo_bots.docs_writer_bot
python -m ai_shell.demo_bots.pylint_bot
python -m ai_shell.demo_bots.test_writer_bot
python -m ai_shell.demo_bots.tool_tester_bot
python -m ai_shell.demo_bots.todo_bot

This is the python interface to the tools, how you're expected to wire up the tool to your bot.

import ai_shell

cat = ai_shell.CatTool(".")
print(cat.cat(["file.py"]))
print(cat.cat_markdown(["file.py"]))

ls = ai_shell.LsTool(".")
print(ls.ls("docs"))
print(ls.ls_markdown("docs"))

This is the smallest example to illustrate basic capabilities, also see here.

import asyncio
import ai_shell


async def main():
    def static_keep_going(toolkit: ai_shell.ToolKit):
        usage = toolkit.get_tool_usage_for("ls")
        if usage["count"] > 0:
            return (
                "Great job! You've used ls. Summarize in paragraph form and we're done."
            )
        return (
            "You haven't used the ls tool yet. Do you have access to the ls tool? If"
            " there is a problem report it to the report_text tool to end the session."
        )

    # Creates temporary bots
    bot = ai_shell.TaskBot(
        ai_shell.Config(),
        name="Folder inspection bot.",
        bot_instructions="Run the ls tool and tell me what you see.",
        model="gpt-4o-mini",
        dialog_logger_md=ai_shell.DialogLoggerWithMarkdown("./tmp"),
    )
    await bot.initialize()
    the_ask = f"""You are in the './' folder. You do not need to guess the pwd, it is './'. 
    Run ls and tell me what you see in paragraph format."""
    await bot.basic_tool_loop(
        the_ask=the_ask,
        root_folder="./src",
        tool_names=[
            "ls",
            "report_text",
        ],
        keep_going_prompt=static_keep_going,
    )


if __name__ == "__main__":
    asyncio.run(main())

This is the cli interface, which is intended for testing, not for bot usage.

ais cat_markdown --file-paths pyproject.toml

Features in Brief

• Many cli-like tools interfaces, such as ls, cat, grep, head, tail, and git.
• OpenAI glue for all cli tools.
• UX with a bot in mind.
• Security with mischievous but not especially malicious bot in mind.
• Bot (Assistant) boilerplate help
• Support for bots doing one shot tool use and goal function driven tool use.
• Bot have extensibility points.
• TODO: plugin system for tools.

Analogues supported today

Directories: ls, find

Files: cat, grep, head, tail

Editing: sed, ed, edlin, patch, replace, insert, rewrite, write new

Data: cut

Other: pycat, token counter, git

Tasking: todo

n.b. Every file is read and written as utf-8 strings.

Prior Art

Running "code interpretor" on your own machine

• open-interpretor - This is the closest to ai_shell I've found.

ai_shell draws inspiration from various command-line interface (CLI) tools and shell environments, integrating features from traditional shells with OpenAI's language models. It is designed to provide an easy and secure interface for AI-assisted file system interactions, keeping in mind both usability and safety.

All of these use the Completions API (last I checked):

• openai-functions
• ActionWeaver
• openai-function-calling
• gptfunctionutils
• openai-decorator
• openai-agent
• openai-func-parser
• openai-function-call
• langjam.func_to_tool
• schemafunc

Uses Assistant/Beta API -openai_assistant_toolkit minimal code in this library.

Documentation

• Features
• Design
• Use Cases
• TODO
• API docs, pdoc3 style