You specify what kind of app you want to build. Then, GPT Pilot asks clarifying questions, creates the product and technical requirements, sets up the environment, and starts coding the app step by step, like in real life while you oversee the development process. It asks you to review each task it finishes or to help when it gets stuck. This way, GPT Pilot acts as a coder while you are a lead dev who reviews code and helps when needed.
- 🔌 Requirements
- 🚦How to start using gpt-pilot?
- 🧑💻️ Other arguments
- 🔎 Examples
- 🏛 Main pillars of GPT Pilot
- 🏗 How GPT Pilot works?
- 🕴How's GPT Pilot different from Smol developer and GPT engineer?
- 🍻 Contributing
- 🔗 Connect with us
The goal of GPT Pilot is to research how much can GPT-4 be utilized to generate fully working, production-ready apps while the developer oversees the implementation.
The main idea is that AI can write most of the code for an app (maybe 95%) but for the rest 5%, a developer is and will be needed until we get full AGI.
I've broken down the idea behind GPT Pilot and how it works in the following blog posts:
[Part 1/3] High-level concepts + GPT Pilot workflow until the coding part
[Part 2/3] GPT Pilot coding workflow
[Part 3/3] Other important concepts and future plans (COMING UP)
gpt_pilot_demo.mp4
- Python 3
- PostgreSQL (optional, projects default is SQLite)
- DB is needed for multiple reasons like continuing app development if you had to stop at any point or app crashed, going back to specific step so that you can change some later steps in development, easier debugging, for future we will add functionality to update project (change some things in existing project or add new features to the project and so on)...
After you have Python and PostgreSQL installed, follow these steps:
git clone https://github.com/Pythagora-io/gpt-pilot.git(clone the repo)cd gpt-pilotpython -m venv pilot-env(create a virtual environment)source pilot-env/bin/activate(or on Windowspilot-env\Scripts\activate) (activate the virtual environment)pip install -r requirements.txt(install the dependencies)cd pilotmv .env.example .env(create the .env file)- Add your environment (OpenAI/Azure), your API key and the SQLite/PostgreSQL database info to the
.envfile- to change from SQLite to PostgreSQL in your .env just set
DATABASE_TYPE=postgres
- to change from SQLite to PostgreSQL in your .env just set
python db_init.py(initialize the database)python main.py(start GPT Pilot)
After, this, you can just follow the instructions in the terminal.
All generated code will be stored in the folder workspace inside the folder named after the app name you enter upon starting the pilot.
IMPORTANT: To run GPT Pilot, you need to have PostgreSQL set up on your machine
git clone https://github.com/Pythagora-io/gpt-pilot.git(clone the repo)- Update the
docker-compose.ymlenvironment variables - run
docker compose build. this will build a gpt-pilot container for you. - run
docker compose up. - access web terminal on
port 7681 python db_init.py(initialize the database)python main.py(start GPT Pilot)
This will start two containers, one being a new image built by the Dockerfile and a postgres database. The new image also has ttyd installed so that you can easily interact with gpt-pilot. Node is also installed on the image and port 3000 is exposed.
If not provided, the ProductOwner will ask for these values
app_type is used as a hint to the LLM as to what kind of architecture, language options and conventions would apply. If not provided, prompts.prompts.ask_for_app_type() will ask for it.
See const.common.ALL_TYPES: 'Web App', 'Script', 'Mobile App', 'Chrome Extension'
Continue working on an existing app using app_id
python main.py app_id=<ID_OF_THE_APP>or workspace path:
python main.py workspace=<PATH_TO_PROJECT_WORKSPACE>Each user can have their own workspace path for each App.
These values will be saved to the User table in the DB.
python main.py user_id=me_at_workIf not specified, user_id defaults to the OS username, but can be provided explicitly if your OS username differs from your GitHub or work username. This value is used to load the App config when the workspace arg is provided.
If not specified email will be parsed from ~/.gitconfig if the file exists.
See also What's the purpose of arguments.password / User.password?
The Architect by default favours certain technologies including:
- Node.JS
- MongoDB
- PeeWee ORM
- Jest & PyUnit
- Bootstrap
- Vanilla JavaScript
- Socket.io
If you have your own preferences, you can have a deeper conversation with the Architect.
python main.py advanced=TrueContinue working on an existing app from a specific step (eg: user_tasks)
python main.py app_id=<ID_OF_THE_APP> step=<STEP_FROM_CONST_COMMON>Continue working on an existing app from a specific development step
python main.py app_id=<ID_OF_THE_APP> skip_until_dev_step=<DEV_STEP>This is basically the same as step but during the actual development process. If you want to play around with gpt-pilot, this is likely the flag you will often use.
Erase all development steps previously done and continue working on an existing app from start of development
python main.py app_id=<ID_OF_THE_APP> skip_until_dev_step=0Here are a couple of example apps GPT Pilot created by itself:
- 💬 Prompt:
A simple chat app with real time communication ▶️ Video of the app creation process- 💻️ GitHub repo
- 💬 Prompt:
Build a simple markdown editor using HTML, CSS, and JavaScript. Allow users to input markdown text and display the formatted output in real-time. ▶️ Video of the app creation process- 💻️ GitHub repo
- 💬 Prompt:
Create a simple timer app using HTML, CSS, and JavaScript that allows users to set a countdown timer and receive an alert when the time is up. ▶️ Video of the app creation process- 💻️ GitHub repo
- For AI to create a fully working app, a developer needs to be involved in the process of app creation. They need to be able to change the code at any moment and GPT Pilot needs to continue working with those changes (e.g. add an API key or fix an issue if an AI gets stuck)
- The app needs to be written step by step as a developer would write it - Let's say you want to create a simple app, and you know everything you need to code and have the entire architecture in your head. Even then, you won't code it out entirely, then run it for the first time and debug all the issues at once. Rather, you will implement something simple, like add routes, run it, see how it works, and then move on to the next task. This way, you can debug issues as they arise. The same should be in the case when AI codes. It will make mistakes for sure so in order for it to have an easier time debugging issues and for the developer to understand what is happening, the AI shouldn't just spit out the entire codebase at once. Rather, the app should be developed step by step just like a developer would code it - e.g. setup routes, add database connection, etc.
- The approach needs to be scalable so that AI can create a production ready app
- Context rewinding - for solving each development task, the context size of the first message to the LLM has to be relatively the same. For example, the context size of the first LLM message while implementing development task #5 has to be more or less the same as the first message while developing task #50. Because of this, the conversation needs to be rewound to the first message upon each task. See the diagram here.
- Recursive conversations are LLM conversations that are set up in a way that they can be used “recursively”. For example, if GPT Pilot detects an error, it needs to debug it but let’s say that, during the debugging process, another error happens. Then, GPT Pilot needs to stop debugging the first issue, fix the second one, and then get back to fixing the first issue. This is a very important concept that, I believe, needs to work to make AI build large and scalable apps by itself. It works by rewinding the context and explaining each error in the recursion separately. Once the deepest level error is fixed, we move up in the recursion and continue fixing that error. We do this until the entire recursion is completed.
- TDD (Test Driven Development) - for GPT Pilot to be able to scale the codebase, it will need to be able to create new code without breaking previously written code. There is no better way to do this than working with TDD methodology. For each code that GPT Pilot writes, it needs to write tests that check if the code works as intended so that whenever new changes are made, all previous tests can be run.
The idea is that AI won't be able to (at least in the near future) create apps from scratch without the developer being involved. That's why we created an interactive tool that generates code but also requires the developer to check each step so that they can understand what's going on and so that the AI can have a better overview of the entire codebase.
Obviously, it still can't create any production-ready app but the general concept of how this could work is there.
Here are the steps GPT Pilot takes to create an app:
- You enter the app name and the description
- Product Owner agent asks a couple of questions to understand the requirements better
- Product Owner agent writes user stories and asks you if they are all correct (this helps it create code later on)
- Architect agent writes up technologies that will be used for the app
- DevOps agent checks if all technologies are installed on the machine and installs them if they are not
- Tech Lead agent writes up development tasks that Developer will need to implement. This is an important part because, for each step, Tech Lead needs to specify how the user (real world developer) can review if the task is done (e.g. open localhost:3000 and do something)
- Developer agent takes each task and writes up what needs to be done to implement it. The description is in human-readable form.
- Finally, Code Monkey agent takes the Developer's description and the existing file and implements the changes into it. We realized this works much better than giving it to Developer right away to implement changes.
For more details on the roles of agents employed by GPT Pilot refer to AGENTS.md
- GPT Pilot works with the developer to create fully working production-ready app - I don't think that AI can (at least in the near future) create apps without a developer being involved. So, GPT Pilot codes the app step by step just like a developer would in real life. This way, it can debug issues as they arise throughout the development process. If it gets stuck, you, the developer in charge, can review the code and fix the issue. Other similar tools give you the entire codebase at once - this way, bugs are much harder to fix both for AI and for you as a developer.
- Works at scale - GPT Pilot isn't meant to create simple apps but rather so it can work at any scale. It has mechanisms that filter out the code so in each LLM conversation, it doesn't need to store the entire codebase in context, but it shows the LLM only the code that is relevant for the current task it's working on. Once an app is finished, you can always continue working on it by writing instructions on what feature you want to add.
If you are interested in contributing to GPT Pilot, I would be more than happy to have you on board but also help you get started. Feel free to ping [email protected] and I'll help you get started.
Since this is a research project, there are many areas that need to be researched on both practical and theoretical levels. We're happy to hear how can the entire GPT Pilot concept be improved. For example, maybe it would work better if we structured functional requirements differently or maybe technical requirements need to be specified in a different way.
Other than the research, GPT Pilot needs to be debugged to work in different scenarios. For example, we realized that the quality of the code generated is very sensitive to the size of the development task. When the task is too broad, the code has too many bugs that are hard to fix but when the development task is too narrow, GPT also seems to struggle in getting the task implemented into the existing code.
🌟 As an open source tool, it would mean the world to us if you starred the GPT-pilot repo 🌟
💬 Join the Discord server to get in touch.