clone-interpreter ✨

a fully replicated OpenAI GPT Code Interpreter sandbox, that is deployable to modal

public web app coming in the next few days

updates on twitter @n_raidenai

Features

• Full control over sandbox environment

  • internet access

  • custom packages installs

  • deployable to modal (which has GPU availability)

• Realtime session persistence

  • user messages

  • generated code / texts

  • generated files

  • can be wrapped in an API

• Custom commands. Current examples :

  • /run and /run_bg for shell commands

    • i.e. /run ls -la

  • /upload for file transfer to sandbox

    • i.e. /upload https://example.com/some.pdf

  • (soon) /doc for fast document interaction, to simulate

    • i.e. /doc https://example.com/paper.pdf what are the main conclusions presented in the abstract

  • /done to close session

How To

note: the web app will be available in a few days, and will abstract away a lot of the configuration steps below

You need :

• a modal account : to deploy the sandbox

• a firebase project with firestore enabled : for chat message subscriptions and session persistence

• a cloud storage bucket : to store and access files generated in the sandbox

Steps :

• installs & service accounts

  • clone the repo

  • install python packages pip install -r requirements.txt

  • configure modal if you haven't done so yet

  • to install chat cli, which is used to interact with the sandbox (temporary solution until web app is done), you need to have nodeJS installed, run :

    • cd cli && npm i

  • create service accounts credentials and save them in the root folder, for :

    • firestore

      • save it as firestoreServiceAccount.json

      • see how to

    • cloud storage

      • save it as cloudStorageServiceAccount.json

      • see how to

• Config & deploy

  • open modal_app.py, configure the variables at the top of the file as you see fit

  • from root folder, run modal deploy modal_app.py. (note: the first build takes a lot of time, i'll push to dockerhub later on to speed this step up. after the first build, it will only take few seconds for each update) after the build, it should output :

    • ✓ Created objects.
      ├── 🔨 Created new_session => https://{MODAL_USER}--clone-interpreter-session-new-session.modal.run
      ✓ App deployed! 🎉
      

  • to configure the sandbox url in the chat cli,

    • open cli/new_session.js and replace the MODAL_APP_URL with the URL your received above:

      • const MODAL_APP_URL = 'https://{MODAL_USER}--clone-interpreter-session-new-session.modal.run/'

    • open cli/watch.js and replace your cloud storage bucket name, to match the one you specified in the modal_app.py config:

      • const BUCKET_NAME = 'your_cloud_storage_bucket_name'

  • that's all

• How to interact

  • Chat CLI :

    • cd cli, then

    • node new_session, it will output 2 node commands, run each in a different shell from the same cli folder:

      • node watch demo_user example01b1ff2670ba9c06ee1b2f90992b88ce

        • observes the session in real time and logs every update

        • downloads the files generated in the sandbox, saved under session-file-sync folder in root

      • node chat https://{MODAL_USER}--clone-interpreter-session-new-session.modal.run/ demo_user example01b1ff2670ba9c06ee1b2f90992b88ce

        • is used to chat with to the sandbox

        • note: i still haven't added queuing and will probably do alongside the web app, so make sure you only send messages when the observer (the watch shell) is inactive

  • Web app :

    • soon

• How to Play

  • Custom commands examples:

    • /run ls -la to run a shell command and wait for output

    • /run_bg some_command some_args to start a shell command in background

    • /upload https://example.com/dataset.csv to have the file inside the sandbox

    • (soon) /doc https://example.com/paper.pdf list the benefits of swimming to do operations on documents

  • To close a session, send /done command in the chat, otherwise the session stays open until timeout and consumes your modal credits uselessly

---

Show what you're building with clone-intepreter, reach out on twitter @n_raidenai