The Dumbest April Fools Joke Ever™

coahGPT was an april fools joke. the whole thing. i built an entire AI chat app with accounts and email verification and a CLI and an openai-compatible API and a whole persona system that talks like me. it ran on a my server with a 3060 for like 2 days. now its open source so you can run it yourself if you want a chat app that says "ngl" and "ts is fire" in every response.

anyway heres how to actually run it yourself.

self-hosting

requirements

• Go 1.22+
• Ollama running locally (or anywhere reachable)
• Node.js 20+ (for building the frontend)

setup

# clone it
git clone https://github.com/coah80/coahgpt.git
cd coahgpt

# pull the model (default is qwen3:8b, change in internal/ollama/client.go if u want)
ollama pull qwen3:8b

# build the frontend
cd web
npm install
npm run build
cd ..

# build and run the server
go build -o coahgpt-server ./cmd/coahgpt-server
./coahgpt-server

server starts on http://localhost:8095. open it in a browser and start chatting.

environment variables

changing the model

edit internal/ollama/client.go and change the Model constant:

const Model = "qwen3:8b" // change this to whatever u want

then rebuild. any model ollama supports works.

how it works

• Go backend serves the API and the static frontend
• SvelteKit frontend (static build) talks to the backend via SSE streaming
• Ollama handles the actual inference
• conversations are stored in your browser's localStorage. nothing is saved server-side
• theres a system prompt in internal/persona/prompt.go that makes it talk like a zoomer. modify or delete it idc
• the /v1/chat/completions endpoint is openai-compatible so you can point other tools at it

project structure

cmd/coahgpt-server/    server entrypoint
internal/
  api/                 http handlers, middleware, rate limiting
  chat/                in-memory session store
  ollama/              ollama client
  persona/             system prompt
  search/              web search (duckduckgo scraping)
web/                   sveltekit frontend (catppuccin mocha theme)

features

• streaming chat with SSE
• web search mode (prefix your message with [Web Search])
• deep research mode (prefix with [Deep Research], fetches more results)
• prompt injection detection (tries to catch jailbreaks)
• rate limiting (60 requests/min per IP)
• openai-compatible API at /v1/chat/completions
• cat that meows when you click it