/llamacoder

Open source Claude Artifacts – built with Llama 3.1 405B

Primary LanguageTypeScript

Llama Coder

An open source Claude Artifacts – generate small apps with one prompt. Powered by Llama 3 405B & Together.ai.

Tech stack

  • Llama 3.1 405B from Meta for the LLM
  • Together AI for LLM inference
  • Sandpack for the code sandbox
  • Next.js app router with Tailwind
  • Helicone for observability
  • Plausible for website analytics

Cloning & running

  1. Clone the repo: git clone https://github.com/Nutlope/llamacoder
  2. Create a .env file and add your Together AI API key: TOGETHER_API_KEY=
  3. Run npm install and npm run dev to install dependencies and run locally

Future Tasks

  • Experiment with a prompt rewriter and launch this as well
  • Make the toast that opens better like a modal for sharability
  • Add sharability to people can take their apps and share them publicly
  • Add the ability to toggle on and off shadcn components and others
  • Launch support for different themes – somehow pass down variables to components
  • Add dynamic OG images to the specific generations & include the prompt
  • Add more dynamic OG images for playwright
  • Address issue of ability to publish the same app repeatedly
  • Try chain of thought reasoning to see if it works better overall
  • Encourage best practices by making the input and textarea & having pills to generate apps w/ good prompts
  • Add more customizability in terms of changing the prompt, temperature, ect...
  • Save previous versions so people can go back and forth between the generated ones
  • Could be nice to show a "featured apps" route on the site on /featured. Have a /id/${prompt} dynamic route that can display a bunch of nice example apps in the sandbox ready to go
  • Support more languages starting with Python, check out E2B
  • Try chain of thought reasoning to see if it works better overall
  • Try finetuning a smaller model on good prompts from 405b or GPT-4/Claude
  • Add dark mode to the site overall, nice design change
  • Surface errors better in codesandbox to the user so people know what is wrong
  • Think about how to have 405B correct itself (sometimes it makes up imports)
  • New route for updateCode that only sends the latest generated code + the modify request
  • Fix bug where if a user edits the code, then does a change, it doesn't use the edited code
  • Add rate limiting with redis upstash if the traffic gets too high
  • Try to add a consistent component library like shadcn
  • Apply code diffs directly instead of asking the model to generate the code from scratch
  • Add the ability to upload things like a screenshot for it to start from that