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
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.
Open.Interpreter.Demo.mp4
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 on 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 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.
ⓘ Issues running locally? Read our new GPU setup guide and Windows setup guide.
You can run interpreter
in local mode from the command line to use Code Llama
:
interpreter --local
Or run any Hugging Face model locally by running --local
in conjunction with a repo ID (e.g. "tiiuae/falcon-180B"):
interpreter --local --model tiiuae/falcon-180B
You can easily modify the max_tokens
and context_window
(in tokens) of locally running models.
Smaller context windows will use less RAM, so we recommend trying a shorter window if GPU is failing.
interpreter --max_tokens 2000 --context_window 16000
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
...
> %debug true <- Turns on debug mode
> %debug 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:
• %debug [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.
• %undo
: Remove previous messages and its 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]
: Calculate the tokens used by the current conversation's messages and estimate their cost, and optionally calculate the tokens and estimated cost of a prompt
if one is provided. Relies on LiteLLM's cost_per_token()
method for estimated cost.
• %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.
- Create a new
config.turbo.yaml
fileinterpreter --config --config_file config.turbo.yaml
- Edit the
config.turbo.yaml
file to setmodel
togpt-3.5-turbo
- Run Open Interpreter with the
config.turbo.yaml
configurationinterpreter --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, Response
import interpreter
app = FastAPI()
@app.get("/chat")
def chat_endpoint(message):
return Response(interpreter.chat(message, stream=True), 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.
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 associated with executing arbitrary code.
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