Let language models run code on your computer.
An open-source, locally running implementation of OpenAI's Code Interpreter.

Get early access to the desktop app‎ ‎ |‎ ‎ Read our new docs

Update: ● 0.1.12 supports --vision (experimental).

pip install open-interpreter

interpreter

Open Interpreter lets LLMs run code (Python, Javascript, Shell, and more) locally. You can chat with Open Interpreter through a ChatGPT-like interface in your terminal by running $ interpreter after installing.

This provides a natural-language interface to your computer's general-purpose capabilities:

• Create and edit photos, videos, PDFs, etc.
• Control a Chrome browser to perform research
• Plot, clean, and analyze large datasets
• ...etc.

⚠️ Note: You'll be asked to approve code before it's run.

pip install open-interpreter

After installation, simply run interpreter:

interpreter

import interpreter

interpreter.chat("Plot AAPL and META's normalized stock prices") # Executes a single command
interpreter.chat() # Starts an interactive chat

OpenAI's release of Code Interpreter with GPT-4 presents a fantastic opportunity to accomplish real-world tasks with ChatGPT.

However, OpenAI's service is hosted, closed-source, and heavily restricted:

• No internet access.
• Limited set of pre-installed packages.
• 100 MB maximum upload, 120.0 second runtime limit.
• State is cleared (along with any generated files or links) when the environment dies.

Open Interpreter overcomes these limitations by running in your local environment. It has full access to the internet, isn't restricted by time or file size, and can utilize any package or library.

This combines the power of GPT-4's Code Interpreter with the flexibility of your local development environment.

Update: The Generator Update (0.1.5) introduced streaming:

message = "What operating system are we on?"

for chunk in interpreter.chat(message, display=False, stream=True):
  print(chunk)

To start an interactive chat in your terminal, either run interpreter from the command line:

interpreter

Or interpreter.chat() from a .py file:

interpreter.chat()

You can also stream each chunk:

message = "What operating system are we on?"

for chunk in interpreter.chat(message, display=False, stream=True):
  print(chunk)

For more precise control, you can pass messages directly to .chat(message):

interpreter.chat("Add subtitles to all videos in /videos.")

# ... Streams output to your terminal, completes task ...

interpreter.chat("These look great but can you make the subtitles bigger?")

# ...

In Python, Open Interpreter remembers conversation history. If you want to start fresh, you can reset it:

interpreter.reset()

interpreter.chat() returns a List of messages, which can be used to resume a conversation with interpreter.messages = messages:

messages = interpreter.chat("My name is Killian.") # Save messages to 'messages'
interpreter.reset() # Reset interpreter ("Killian" will be forgotten)

interpreter.messages = messages # Resume chat from 'messages' ("Killian" will be remembered)

You can inspect and configure Open Interpreter's system message to extend its functionality, modify permissions, or give it more context.

interpreter.system_message  = """
Run shell commands with -y so the user doesn't have to confirm them.
"""
print(interpreter.system_message)

Open Interpreter uses LiteLLM to connect to hosted language models.

You can change the model by setting the model parameter:

interpreter --model gpt-3.5-turbo
interpreter --model claude-2
interpreter --model command-nightly

In Python, set the model on the object:

interpreter.model = "gpt-3.5-turbo"

Find the appropriate "model" string for your language model here.

Open Interpreter uses LM Studio to connect to local language models (experimental).

Simply run interpreter in local mode from the command line:

interpreter --local

You will need to run LM Studio in the background.

1. Download https://lmstudio.ai/ then start it.
2. Select a model then click ↓ Download.
3. Click the ↔️ button on the left (below 💬).
4. Select your model at the top, then click Start Server.

Once the server is running, you can begin your conversation with Open Interpreter.

(When you run the command interpreter --local, the steps above will be displayed.)

Note: Local mode sets your context_window to 3000, and your max_tokens to 1000. If your model has different requirements, set these parameters manually (see below).

You can modify the max_tokens and context_window (in tokens) of locally running models.

For local mode, smaller context windows will use less RAM, so we recommend trying a much shorter window (~1000) if it's is failing / if it's slow. Make sure max_tokens is less than context_window.

interpreter --local --max_tokens 1000 --context_window 3000

To help contributors inspect Open Interpreter, --debug mode is highly verbose.

You can activate debug mode by using it's flag (interpreter --debug), or mid-chat:

$ interpreter
...
> �bug true <- Turns on debug mode

> �bug false <- Turns off debug mode

In the interactive mode, you can use the below commands to enhance your experience. Here's a list of available commands:

Available Commands:

• �bug [true/false]: Toggle debug mode. Without arguments or with true it enters debug mode. With false it exits debug mode.
• %reset: Resets the current session's conversation.
• %undo: Removes the previous user message and the AI's response from the message history.
• %save_message [path]: Saves messages to a specified JSON path. If no path is provided, it defaults to messages.json.
• %load_message [path]: Loads messages from a specified JSON path. If no path is provided, it defaults to messages.json.
• %tokens [prompt]: (Experimental) Calculate the tokens that will be sent with the next prompt as context and estimate their cost. Optionally calculate the tokens and estimated cost of a prompt if one is provided. Relies on LiteLLM's cost_per_token() method for estimated costs.
• %help: Show the help message.

Open Interpreter allows you to set default behaviors using a config.yaml file.

This provides a flexible way to configure the interpreter without changing command-line arguments every time.

Run the following command to open the configuration file:

interpreter --config

Open Interpreter supports multiple config.yaml files, allowing you to easily switch between configurations via the --config_file argument.

Note: --config_file accepts either a file name or a file path. File names will use the default configuration directory, while file paths will use the specified path.

To create or edit a new configuration, run:

interpreter --config --config_file $config_path

To have Open Interpreter load a specific configuration file run:

interpreter --config_file $config_path

Note: Replace $config_path with the name of or path to your configuration file.

1. Create a new config.turbo.yaml file

   interpreter --config --config_file config.turbo.yaml
   

2. Edit the config.turbo.yaml file to set model to gpt-3.5-turbo
3. Run Open Interpreter with the config.turbo.yaml configuration

   interpreter --config_file config.turbo.yaml
   

You can also load configuration files when calling Open Interpreter from Python scripts:

import os
import interpreter

currentPath = os.path.dirname(os.path.abspath(__file__))
config_path=os.path.join(currentPath, './config.test.yaml')

interpreter.extend_config(config_path=config_path)

message = "What operating system are we on?"

for chunk in interpreter.chat(message, display=False, stream=True):
  print(chunk)

The generator update enables Open Interpreter to be controlled via HTTP REST endpoints:

# server.py

from fastapi import FastAPI
from fastapi.responses import StreamingResponse
import interpreter

app = FastAPI()

@app.get("/chat")
def chat_endpoint(message: str):
    def event_stream():
        for result in interpreter.chat(message, stream=True):
            yield f"data: {result}\n\n"

    return StreamingResponse(event_stream(), media_type="text/event-stream")

@app.get("/history")
def history_endpoint():
    return interpreter.messages

pip install fastapi uvicorn
uvicorn server:app --reload

Since generated code is executed in your local environment, it can interact with your files and system settings, potentially leading to unexpected outcomes like data loss or security risks.

⚠️ Open Interpreter will ask for user confirmation before executing code.

You can run interpreter -y or set interpreter.auto_run = True to bypass this confirmation, in which case:

• Be cautious when requesting commands that modify files or system settings.
• Watch Open Interpreter like a self-driving car, and be prepared to end the process by closing your terminal.
• Consider running Open Interpreter in a restricted environment like Google Colab or Replit. These environments are more isolated, reducing the risks of executing arbitrary code.

There is experimental support for a safe mode to help mitigate some risks.

Open Interpreter equips a function-calling language model with an exec() function, which accepts a language (like "Python" or "JavaScript") and code to run.

We then stream the model's messages, code, and your system's outputs to the terminal as Markdown.

Thank you for your interest in contributing! We welcome involvement from the community.

Please see our Contributing Guidelines for more details on how to get involved.

Open Interpreter is licensed under the MIT License. You are permitted to use, copy, modify, distribute, sublicense, and sell copies of the software.

Note: This software is not affiliated with OpenAI.

Having access to a junior programmer working at the speed of your fingertips ... can make new workflows effortless and efficient, as well as open the benefits of programming to new audiences.

— OpenAI's Code Interpreter Release