GitHub - andi0b/vibe-todo: A todo app built entirely in bash with microservices. Because we could.

vibe-todo

Because the world needed another todo app. But in bash. With microservices.

What is this?

This is a fully-featured todo list application built entirely in bash scripts using netcat as the HTTP server. Yes, you read that right. No Node.js. No Python. No frameworks. Just pure, unadulterated shell scripting and questionable life choices.

Someone asked an AI to vibe-code a todo app and it chose violence.

Architecture

┌─────────┐     ┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│ Browser │────▶│ API Gateway:8000│────▶│ Todo Service:8002│────▶│Storage Svc:8001 │
└─────────┘     └─────────────────┘     └──────────────────┘     └─────────────────┘
                        │                        │                       │
                        ▼                        ▼                       ▼
                ┌─────────────────┐     ┌────────────────┐       ┌─────────────┐
                │Frontend Svc:8003│     │ Bashis:6379    │       │ todos.json  │
                └─────────────────┘     └────────────────┘       └─────────────┘
                        │
                        ▼
                ┌─────────────────┐
                │ LLM Svc:8004    │  ← A transformer. In AWK. Yes, really.
                └─────────────────┘

Six microservices. For a todo list. One of them is a neural network. Running on netcat. Because obviously.

Bashis: A Redis Clone in Bash

We would've called it Redish™ but Redis didn't grant us that. So instead, meet Bashis - a Redis-compatible key-value cache server written entirely in bash. It speaks actual RESP protocol and works with real redis-cli. Because writing our own caching layer was easier than adding a dependency.

LLM Service: A Transformer in AWK (yes, AWK)

We implemented a GPT-style transformer from scratch. First in bash, then rewrote it in AWK because we asked "can we go faster?" instead of "should we stop?". Fixed-point arithmetic and Taylor series approximations for activation functions. It generates text at approximately one token every few seconds. This is art.

What it does:

• Full transformer architecture (embeddings, attention, FFN, layer norm)
• Trained on Shakespeare (or whatever text you provide)
• Generates text that is... creative

What we might use it for:

• We genuinely don't know yet
• Maybe AI-powered todo suggestions?
• Maybe it just sits there, generating Shakespeare
• The point is it exists

Does it work?

• Yes, verified against Python reference implementation
• Output quality depends on your patience for training
• It will definitely generate something

See llm-architecture.md for the full technical deep-dive.

Features

• Add todos (revolutionary)
• Mark todos as complete (groundbreaking)
• Delete todos (disruptive innovation)
• A beautiful purple gradient UI (the only thing that makes sense here)
• Redis-compatible caching layer written in bash (Bashis) - works with actual redis-cli
• LLM inference engine in AWK - a GPT-style transformer that runs at geological speeds
• File locking with flock because race conditions in your bash todo app would be embarrassing
• JSON parsing with sed and grep because who needs jq
• Matrix multiplication in bash arrays because linear algebra doesn't require C

Prerequisites

• Bash (the only dependency that matters)
• nc (netcat) - your enterprise-grade HTTP server
• Docker (optional, for running in containers without Dockerfiles)
• A complete disregard for conventional software engineering practices
• The audacity

Running

Start the microservices empire:

cd todo-microservices
./start.sh

Visit http://localhost:8000 and witness the glory.

Running with Docker

For those who want containers but refuse to write Dockerfiles:

./docker-build.sh           # Build images using docker create/commit (no Dockerfiles)
./docker-compose.sh up -d   # Start the stack (YAML not invited)
./docker-compose.sh ps      # Admire your orchestration
./docker-compose.sh down    # Tear it all down

Your Docker Desktop won't suspect a thing - it sees a respectable compose project.

Stopping

This kills the dream (and processes on ports 8000-8004 and 6379).

Testing

Yes, there are tests. We're not complete animals.

Technical Highlights

• HTTP/1.1 implementation from scratch - because importing an HTTP library is for the weak
• RESP protocol implementation - Bashis speaks Redis wire protocol, tested with real redis-cli
• GPT-style transformer in AWK - with fixed-point math, Taylor series, and Newton's method
• Bash coprocesses - bet you didn't know bash could do that
• Proper CORS headers - we're chaotic, not incompetent
• JSON manipulation with regex - just as God intended (now handles colons in values!)
• Service-to-service communication - over HTTP, via netcat, in bash, on localhost
• Caching with associative arrays - declare -A is basically Redis if you squint
• Matrix multiplication in nested loops - O(n³) in AWK and proud of it
• Softmax via Taylor series - because importing numpy is cheating
• Process group management - proper Ctrl+C handling because we're professionals

FAQ

Q: Why? A: Why not?

Q: Should I use this in production? A: Absolutely. Scale it to millions of users. What could go wrong?

Q: Is this a joke? A: This is art.

Q: How do I add authentication? A: Write another bash script. That's how we solve all problems here.

Q: My company wants to adopt this. What's the enterprise licensing? A: Please seek professional help.

Contributing

If you want to add more microservices to this todo app, you might be exactly the kind of person we're looking for. Or avoiding. We're not sure yet.

License

MIT - because even chaos deserves freedom.

Built with vibes, netcat, and an AI that was asked to "just make a todo app" and dramatically misunderstood the assignment.