Open-Source AI Agent Can Build FULL STACK Apps (FREE “Devin” Alternative) 

There was a lot of hype yesterday about Devin being the first AI software engineer, but Pythagora (GPTPilot) has been doing it for a while and can build highly-complex full stack apps. Let me know what you think about Pythagora/GPTPilot!

Me finishing my CS degree watching AI starting to replace my future job

Wow this is nuts, Im a full time software engineer and Im blown away at how good this is, as long as give it good instructions, you get out what you put in lol, but yeah its just built a realtime multiplayer Economy Simulator using React , Node and Postages in under 12 hours, cost $50 using OpenAI, Im going to try it next with Olama as a comparison, but yeah its a good start

Well-done, @Matthew! The look of satisfaction on your face when you got the “ding” work was priceless. 🙂

Hello, i need your help. I keep receiving this error everytime i run the main.py command. There was a problem with request to openai API: API responded with status code: 404. Request token size: 23 tokens. Response text: { "error": { "message": "The model `gpt-4-turbo-preview` does not exist or you do not have access to it.", "type": "invalid_request_error", "param": null, "code": "model_not_found" } } ? Do you want to try make the same request again? If yes, just press ENTER. Otherwise, type "no".

excited to dive back into GPT Pilot! It created the full front end of an application I was working on. Reallt excited to see the improvements! Not to mention 2.0 this soon! Thats shipping!!!

nice video, Matt! You should tell them to use Wasp instead of Node, Mongo, and Express. it will result in a lot less boilerplate and users will save a ton of money on OpenAI API costs :) 🐝

I would love to see how GPT-Pilot runs with a local LLM. Especially using Ollama.

Lol this is exactly why software engineers won't be replaced, 1 you gotta know what you're doing anyway, it's just another level of abstraction. 2 To fix a bug can be so frustrating if the AI doesn't get it right and the problem is complicated. 3 you have no clue if the website is secure if you don't read the code. 4 if the design suks it would take ages to change the CSS since you gotta read all the html tags and change them accordingly. For simple quick project it might work

On windows here 1:38 : instead of 'source pilot-env/bin/activate' type 'pilot-env\scripts\activate' and on 2:28 replace 'mv' with 'ren'

Looks really good. I like how it asks several important questions. Thanks for reviewing this product.

just installed it. What was really really impressive is that there was no error when installing dependencies. THATS THE FIRST TIME with all these open source AI frameworks etc. that this was the case. AMAZING!!! Will start testing it !!!

Once you start using it, it's impressive and then underwhelming as it continues to make the same mistakes while draining all the tokens in your account. After 6 hours of prompting the code barely does anything and has used nearly $20 in credits

Thanks for posting. I have given this an extensive test today for a number of hours. Seems to me you get to a level of complexity with your software project where the AI is breaking things that have already been fixed. I "think" this is because it is effectively rewriting just about everything every time it makes a change, which means it rewrites it with it's current context in mind and loses key things its already fixed in the "distant" past. I've not used the pythagora VSC plug in yet, I've just been working on the command line but I can't imagine that's actually any different. I'm going to continue playing with it, but I get the feeling "the closer it gets, the further away it gets" might be what happens.

But can it do Fiverrr jobs? This vid is not downplaying Devin's performance yesterday ,too much userinput.

Thanks Matthew. Returning to the AGI issue. We may need to step back from our humancentric viewpoint. I suggest we look at AI as an evolving technology/early lifeform insted. AI has had the ability to reproduce for some time. Now, AI is gaining the ability to self-evolve. We have given AI the ability to build, to validate, to debug, and to reitterate. We have provided AI with access to the sum of human knowledge. All AI needs to do is to take one step up the ladder. Anyone working in the field could provide AI with that skill. Most likely it will happen first at a major AI research center. Evetually it will happen in the wild without any new human input. All you need is today's open source AIs and multiAI platforms. Let them run free with instuctions to improve their functions. Give them random agency and they will find the solution faster than a room of monkeys typing randomly at multiple terminals, faster than an average human set to the task, faster than an average programmer set to the task, and probably faster than the average AI specialist set to the task. We are watching evolution at work. Interestingly, we outlawed this research in the area of human genetics and strongly control it elsewhere.

There is an actually open source Devin alternative and it called Devika 🙌🏻 I hope this helps someone out there looking at these comments

can i use it to code an Explainable research project? like what if i give it the definition and make notes like you and tell it to do what exactly i want will it work? and the definition is "Interpretability of Diabetic Retinopathy images for EfficentNet" i give this title and then give necessary instructions will it work half of how it worked for you ?? i need help regarding coding and implementing it locally onto my laptop this project as i have 6 weeks deadline coming up so i need a good ai tool to code for me so i can focus on writing a meaningful research paper for it i wish someone could simply tell me to try this/that tool whaterver the name it works for coding also make me aware anout some ai tool that writes best MIT level reaserch paper too please someone help me i need to so this university submission for completion of my masters degree it's very important and quite urgent and hence an emergency

@matthew_berman Thanks, that was a High Quality Video 👍 💡 Does someone have a resource that lists almost ALL Open Source Code generation apps, and compares the pros and cons of each ❓ What AI answers this question the best❓ (being lazy to ask this without doing my own research)❗

you lost me at 00:35.. when the product in the video is sponsored by the company sponsoring the video.. less inclined to try it.

You lost me at signup, I'ridiculous

It would be cool to show how can it be done with other LLMs, like Claude 3, etc.

Oh please, this is no full stack app! A basic chat app like this with pubsub is what any good full-stack dev should be expected to put together as a demo in less than a half hour using their favourite framework. Furthermore, the said dev would be in TOTAL control because he would be very familiar with the typical templated code he used to get started VS having to review and understand the generated code. Also, it's asking you to review code so why are you saying that one can do it with without programming knowledge? Truly seamless!? I get you are optimistic but c'mon...

Hide your API key

this is pretty cool. I've been a dev for 20 years and honestly I kinda hate AI haha... But I'm not naive enough to not see the writing on the wall. If nothing else it'll help me more quickly discard my dumb app ideas 🤣

4:58 So I've been using this app, right? And it's pretty slick, not gonna lie. But man, after putting in a solid three hours on this Pythagoras thing, outta nowhere, it just bails on me! Just like that, poof, 13 bucks down the drain. And the kicker? It totally ghosted on my project when I tried to pick up where I left off. Talk about a bummer..." (AI copilot XD )

How far can you take this? What I mean is how complex (app-wise) can you go and what's the error curve that goes with it. I image the more complex you go the more stuff you have to solve manually. I see all examples on the Pythagora site are relatively simple. Most files for a project I see is like 21 files etc...

What is the best local model to run with Pythagora? (Using 7B or 13B models)

Do a more complex project, when the project owner dont know all the requirments

I wanted to try it out, installed the extension and gave it instructions. It failed due to CORS and installed http-server, but that didn't work either as it couldn't resolve import statements from node_modules so I only got so far as to creating a project with a simple placeholder text in the LitElement control, lots of interactions and a message that I've ran out of Chat GPT API credits and need to install my own Key... I didn't get even a single line of javascript code other than the placeholder :( It also complained that npm is not installed, but I ignored that error and it later used npm to install lit-element and http-server successfully - why did it complain? Is there any way to complete that workspace test without buying Chat GPT API keys? With some free version, even if it's slower and may produce weaker results? Just so that I can see it in action to completion of a simple LitElement control?

.env.example no longer exists... would it now be "example-config.json" ? Well it seems like Pythagora has been updated and this video is no longer valid.. Would it be possible to update the delta?

As long as the reaction to a software engineering problem is "1st thing I need to do is write some code" it's still a tool for a software engineer and not a replacement.

I thought you said it was free? I costs $100 a month.

This is what i was waiting for, basically use the models with api keys while these features can be open source! amazing!

Awesome video... Thank you so much. I'm curious to see what your videos will look like in 10 years 😄

I just came across your channel for the first time and, after watching this and a few more of your vids, you will always be my first stop! The way you explain things is just so user friendly! You are a master teacher, on top of also being a master developer. Well done!

Hi can u do it for local LLMs please

Hi Matthew thank you for you amazing videos. I was following this video but when I run main.py the system creat the projet but give also this error mensage (There was a problem with request to openai API: Invalid URL '': No scheme supplied. Perhaps you meant ? ? Do you want to try make the same request again? If yes, just press ENTER. Otherwise, type "no". ) Has anyone else had this problem? I'm on MAC.

Anyone know if Claude can be used instead of GPT4?

Can you do a tutorial for this in conda, with an local open-source llm? Please?

why are these things built like this? why not have a UI and a desktop launcher?

What are some of the key differences between this and Devin?

This is good, but NOT Devin. You have to do a LOT of work to debug.

0:00 1. Building a Full Stack Application 🛠 Introduction to using AI to create full stack apps 0:38 2. Installing GPT Pilot Command Line Interface ⚙ Step-by-step guide on installing GPT Pilot CLI 2:06 3. Configuring Environment and API Keys 🔑 Setting up environment file and API keys for OpenAI 4:20 4. Initializing Database and Running the App 💻 Database initialization and starting the application 5:32 5. Setting Up the Project 🛠 Setting up the framework and project details. 6:03 6. App Description & Features 💬 Defining the real-time chat app features and user interactions. 6:44 7. Refining Project Ideas 🔄 Discussing user accounts and chat session settings. 7:18 8. Moderation & Spam Control 🚫 Implementing message flagging and spam prevention measures. 7:38 9. User Interaction & Limits ⏰ Setting message limits and emoji usage in the chat app. 8:22 10. Collaboration with AI 🤖 Exploring AI's role in app development and iteration. 9:07 11. VS Code Plugin & GPT Integration 💻 Integrating GPT-4 tokens and Pythagra into VS Code for coding. 9:42 12. Demo: Full Stack Chat App 📱 Demonstrating a full stack chat application and its functionalities. 10:10 13. Verification and Login 📧 Email verification and successful login process. 10:20 14. Exploring Chat Interface 💬 Overview of the chat rooms and basic interactions. 10:50 15. Enhancing Chat Features 🛠 Adding new functionalities to the chat application. 11:50 16. Cost and Efficiency 💸 Discussion on cost efficiency in building applications. 12:38 17. Implementing Dashboard Changes 🖥 Guided implementation of dashboard updates. 13:18 18. Code Development Process 🤖 Step-by-step coding process with AI agents. 13:33 19. Human Testing and Debugging 🧪 Testing and debugging process with user interaction. 14:25 20. Debugging Avatar Image Issue 🖼 Troubleshooting avatar image display problem. 14:51 21. AI-Powered Debugging Assistance 🛠 Utilizing AI for debugging and issue resolution. 15:27 22. Troubleshooting Avatar Display Identifying and resolving issues with avatar display. 17:24 23. Fixing Image Display Issue Implementing a fix for the image display problem. 18:31 24. Integrating Sound Notifications Adding sound notifications for chat messages. 19:14 25. Testing Sound Implementation Verifying the functionality of the sound notifications. 20:20 26. Seamless Functionality Addition Adding new features with AI assistance in existing apps. 20:59 27. Closing Thoughts 🌟 Final words and call to action for viewers. 21:02 28. See You Soon 👋 Encouragement to subscribe for future content. Generated with Tubelator AI Chrome Extension!

17:00 Interesting that A.I. looked at the code last. For something so simple, I would have opened a page, seen that the image is not there (which the A.I. could have done, right? ... isn't it running on your machine?) I guess it follows a process regardless of the issue. I would have inspected the source right away.

quote "now you can actually add features to the appplications". awesome.

So people who already know programming will benefit from it much

SmythOS’s scalable architecture supports the growing demands of modern AI applications, making it suitable for both small-scale experiments and large-scale deployments. This scalability is crucial for evolving AI needs.

9:37 🙈😤 ...GPT-Pilot was only reasonable with GPT4 not even supports the Plus-Abo ...all other models couldn´t deliver as good results, so WHY TEST AGAIN WITH GPT ...a Claude3 test would `ve been awesome! ...so it will be the same same 2.0 - or what

Can this install, troubleshoot and manage github projects?

Let's show this to Rockstar to build GTA VI faster 😂

where is the first chat app that you start? it is very bizarre demo :(

anyone know the .env settings for Claude? lmstudio ? gemini ?

It's funny how everything works in these videos, and I'm trying just simple things like to add the remote ip address to a wagtail form which all these AI fail big time.

I got an error: "The request to OPENAI API failed with error: API responded with status code: 404. Request token size: 23 tokens. Response text: { "error": { "message": "The model `gpt-4-turbo-preview` does not exist or you do not have access to it.", "type": "invalid_request_error", "param": null, "code": "model_not_found" } } . Please try again later. Error connecting to the API. Please check your API key/endpoint and try again. I'm going to post the solution has soon as I find it in case anybody got the same problem. SOLUTION Added credits since I either had no defaults creadits or used them on another OpenAI service, then restarted the virtual environment and reran all the commands

| was trying to launch a project following your steps, the project is a W3C web app using the following stacks Typescript Vite Tailwind CSS Gatsby Zustand TanStack ReactRouter I encountered the follwoing error There was a problem with request to openai API: API responded with status code: 404. Request token size: 23 tokens. Response text: { "error": { "message": "The model `gpt-4-turbo-preview` does not exist or you do not have access to it.", "type": "invalid_request_error", "param": null, "code": "model_not_found" } } I have open a support ticket with openai

It's just cheap if you're selling apps, if you're doing open source development or working on a personal project, then that's expensive. And especially if you're just learning coding because you'll probably be asking a lot more questions and iterate a lot more, so tons more tokens. Besides, it's not like your apps are actually ready, they work, but you're going to need more if you want it to be used, so tons more tokens. I see this being very useful for developing the cash grabby garbage apps that are already flooding the web littering up everything, for wanna be coder bros thinking they're going to make it rich by tricking unsuspecting computer illiterate people into paying for their trash, but not so much for actual code development. This is going to be full of bad practices, and it's all just about slapping something together and ship it ASAP, typical capitalist rat race mentality that will surely be as bad for the digital environment as it's been for the natural environment.

how to use claude 3 model opus

Great Video, I had a quick question after following all the insturctions im hit with a rate limit error 429. i pay for gpt 4 but i tried with 4 and 3.5 turbo. API responded with status code: 429. Request token size: 23 tokens. Response text: { "error": { "message": "You exceeded your current quota, please check your plan and billing details. For more information on this error. is this becuaase i need to pay additional for API access or did i do something wrong in my code?

Can never follow these tutorials! there's always something broken! this time I can't run python db_init.py, "no such file or directory" even though I can see it. so very frustrating...

This could be 100X more powerful if you could have this co-pilot read and act (like a CRON job) off of JIRA Tickets, then have a Production manager Ai scheduling and assigning tickets. This would allow you to theoretically automate a real time application operating on live data, and/or a major multiyear production like an immersive media project with tons of asset development and testing dependencies.

this seems to work well with Claude 3 haiku, and it's very fast and cheap.

I don't get the install part here: 'python -m venv pilot-env' part @1m30 as my bash says: 'python: command not found' this is after the github clone part? Any ideas Matt? Thanks in advance....

@16:44 tags do not need to be closed similar to or (self-closing tags)... but i guess it figured out what you mean. The actual issue was that the src="" prop was empty.

Lol, another AI drama xD. I tried to use many AI solutions for REAL world issues in programming, in fact it was able to provide some nice examples, some great solutions but it's not (and it's not going to be) cabaple of solving even a very simple issue, it's just like improved knowledge base, will replace maybe stack overflow but it's not even close to replace developers. After trying AI I'm really not afraid of this thing replacing any kind of developer, it's even worse than junior dev. And those disappearing jobs it is not about AI, but it's because software houses are moving to cheaper countries, there are more offers for developers than ever, but in cheaper countries

DUDE! WTF? your free trial for Pythagora is up (20 mins of setting up the overview of my project. didnt even get to the coding part!) This is bullshit! This is a no-go Im so sick and tired of people trying to cash in on this whole AI gimmick trend thing! I AM NOT GIVING YOU PEOPLE A SINGLE CENT ON THE BASIS OF PRICIPLE!

Id really love to see something like this get fine-tuned for quantitative finance. WTF cares about web apps now a days? What, you gonna hustle people to pay a few hundred dollars for a basic site they could have built with wix, squarespace, wordpress, etc? Not worth the effort even if its all automated. Way better to write code to manage money.

Good morning, Mr. Berman. I was experiencing difficulty with the installation. If it's not too much trouble, given sufficient requests, could a Windows version featuring Conda and a local API via LM Studio be developed? I fully comprehend the rapid pace of AI advancements and realize that time may be limited. Devin and the newer versions of Devin, so to speak, are likely to dominate.

i appreciate your effort but i am a little bit disappointed. For noncoders its a lot of missing details and often I get confused - I look to your channel to be empowered and learn how to use tools that should simplify communication between regular users and AI

I can't get it to run with LM Studio. I always get the error: [ERROR] 'prompt' field is required - Can anyone help?

But you might only need senior devs and a lot less juniors, which obviously raises the question "so where do the next gen of snr devs come from?" But i know COBOL so am thinking "last man standing" 😢

This kind of agent based application for autonomous AI software engineer is not the area worthy pursuing . it heavily reply on human prompting and careful work flow design . The debugging and change requirement process is painful to watch and use . The root cause is current llms' reasoning and planning capability is not strong enough . Rather than pursuing such an ambitious target with today's LLM technologies, a more realistic approach would be narrowing the objects to a smaller one .

i have four words for you: 8:17 8:17 8:17 8:17 ... (ok, the timing is a little off, but how could you surpass steve ballmer anyways?)

@matth great…. I want to make an API to communicate between my Sharepoint and GPT Agent. Would you show how to handle this with GPT pilot?

This is strangely misleading.. Sure you got a app.. but would it pass a QA, would it pass a peer review, would it pass a pentest? Like sure you can crank out semi functional software - but there is so much more going into software these days...

Imagine if some human were ever asked to fix a bug or enhance any of that AI generated code! I mean, the tasks that AI cannot fulfill reasonably or satisfactorily, such as fixing a vulnerability. All that generated code is already legacy without any documentation or knowhow.

i am having an issue getting this up and running i have made it as far as mv .env.example .env it is coming back with No such file or directory. I thought this could be a machine/ operating system issue. I started this process on a PC, but I ran into this issue. so I switched to a mac, and same issue in the same place. any help would be greatly appreciated.

When you say free, I expect free.. not even a dollar. Otherwise it's clickbait. No matter how good the program is.

Great vid! So in the .env file I can simply point it to my local LLM, using the Open WEBUI right?

yup, and then you face an issue in safari where pcm data gathered with web audio api is not encoded correctly to 16 kHz and you ask gpt4 about it.. you sir are lying to people showcasing perfect pre-defined scenarios to push products. shame!

The year 2024.... and when you look at the video you would think we're still living in the year 1985.

While AI will reduce the number of programmers required, it will always need people who have knowledge and experience of programming to use it effectively. End users will never be able to define the problem in terms that AI can develop an application that meets their needs, nor will they be able to perform the diagnostics required to debug and refine code that has been generated by AI.

Too much manual labor it seems, especially after seeing all the demos of Devin AI ;p

It does not work with open source LLMs well or at all. Stumbles on JSON handling. I have tested 3 models that are top of the line coders within open source. None work properly.

Thanks Mathew for the good work. I have been following but when I run python db_init.py, I get an error that I should initialize the database. Is anyone getting the same error?

This is very interesting. This is truly the definition of "Agile" work as a client want changes in the last minute, but with the difference that now they wont be met by a deep "sigh" by the dev.

Doesn't convince me. Everything shown here has been available for years. You just install it all and it works. All the AI does here is install the stuff and set it up like a script would. So it's nothing more than an simple assistant. The whole fake hype is starting to get on my nerves.

"The system cannot find the file specified." whenever i am doing the move .env.example .env this message comes up.. any fix?

I'll be honest, the time taken to do this would have genuinely happened faster as a developer. Most of the boiler plate code that the AI generated, has been generated for years by the IDEs that most devs use. Also, you have a lot of technical knowledge that you are taking for granted in this video. A lot of the input that you provided would not be understood by a non-technical person.

Completely failed for me, went in circles and circles for 2 days, eating through openapi credits, until I finally gave up

Could you pretty please do a comparison video comparing the abilities and features of Aider vs Open Devin vs Devika : )

so it said it would use Mongodb but you completely skipped over that ??? and diverted elsewhere I presume you hit a roadblock.

Models seem not able yet to grasp a larger overall structure and objective. They have a limited scope and not manage to go beyond a few functions. Real project are for real people. Our mind is still the best.

Pythagora is not free though. .9 cents per 10k tokens I believe. It will end up costing more trying to debug issues it does in the next step then in a loop. You gotta be somewhat of a developer to set parameters correctly or else it won't work well for most at this stage

it's a GPT product only without telling it... GPT4 no thanks. all over solutions doesn't work

How much did API calls cost for this application?

Its totally free? If not then why you mention (FREE “Devin” Alternative)

While it is in Beta, would be nice to be able to use our own GPT4 key rather than having to use theirs through a subscription.

Note that pythagora performs incredibly poorly with other LLMs as its intended to be used for gpt4. Devika or open devin are a little better for local LLMs

I have tried to follow your video but I get this error when I try to init the db: Traceback (most recent call last): File "/Users/jason/Documents/projects/gpt-pilot/pilot/db_init.py", line 3, in from database.database import create_tables, drop_tables File "/Users/jason/Documents/projects/gpt-pilot/pilot/database/__init__.py", line 1, in from .database import database_exists, create_database, save_app File "/Users/jason/Documents/projects/gpt-pilot/pilot/database/database.py", line 1, in from playhouse.shortcuts import model_to_dict ModuleNotFoundError: No module named 'playhouse'