The Optimization Obsession
Why we rebuilt everything from scratch instead of
bolting AI onto someone else's architecture.
The IDE Market Has a Forking Problem
Look at the landscape of AI coding tools. What do you see?
VSCode fork. VSCode fork. VSCode extension. VSCode fork.
CLI tool. CLI tool. Terminal UI. CLI tool.
VSCode fork with extra steps.
Every tool starts from the same place: take something that already exists, add AI features, ship fast, iterate later. It's a reasonable strategy. It's also why every tool feels the same, performs the same, and hits the same walls.
We're not interested in reasonable strategies.
What If We Started From Zero?
Vibe Mode began with a question: What would an IDE look like if AI wasn't a feature, but the foundation?
Not “code editor + AI sidebar.”
Not “terminal + chat interface.”
Not “fork of X with Y bolted on.”
An IDE where the inference engine and the editor share memory. Where context indexing happens at filesystem speed because it's written in Zig, not JavaScript. Where the terminal emulator doesn't spawn a new process because it's compiled into the same binary.
This is what we mean by “built from scratch.”
The Seven Vibes of Obsessive Engineering
(We're only half-joking about the vibes thing)
Every Byte Earns Its Place
100MB for a complete IDE with inference, terminal, and 15+ provider integrations. If something doesn't justify its size, it doesn't ship. Bloat is a bug.
Memory Is Sacred
Under 100MB at idle. Your RAM exists to run models, not to keep an Electron wrapper warm. We measure memory usage religiously and treat regressions as emergencies.
Latency Is Lying
When your IDE lags, it's lying to you about its capabilities. We target 60fps for all animations, sub-second cold start, and 3-20ms time to first token. Smooth isn't a nice-to-have.
Dependencies Are Debt
Every external dependency is a liability—maintenance burden, security surface, potential bloat. We wrote our own inference engine, our own templating engine, our own everything. When something isn't good enough, we build it ourselves.
Context Is King
The difference between a helpful AI and an annoying one is context. We obsess over giving models the right information: project structure, dependencies, recent changes, coding patterns. An AI that knows your codebase is worth ten that don't.
Native or Nothing
macOS has beautiful APIs. Apple Silicon has incredible performance. Why would we wrap everything in a web browser? SwiftUI for the interface. Metal for the GPU. Zig for the performance-critical paths. Native all the way down.
The Vibes Must Flow
Yes, we named it "Vibe Mode." Yes, we're aware that's ridiculous. But here's the thing: when your tools disappear and you're just creating, that's a vibe. We're engineering that feeling—the flow state where great work happens. Call it vibes. We do.
This Is What Obsession Gets You
When we said “built from scratch,” we meant it:
- Inference engine (VibeMLX) — written in Zig
- Templating engine (vibe-jinja) — written in Zig
- Context management (Samsara) — written in Zig
- HTTP/WebSocket server — written in Zig
- Terminal integration (Ghostty) — integrated, not spawned
- File watching & indexing — native APIs, not polling
- UI framework — SwiftUI, not Electron
The result: A 100MB download that outperforms 1GB+ competitors in every metric that matters for development flow.