The Stack That Makes the Vibes Possible
Every component built for one purpose:
getting out of your way while making you faster.
Inference Runtime
VibeMLX
17MB of Pure Inference
Most local inference solutions require Python environments, pip packages, model conversion pipelines, and prayers to the dependency gods. VibeMLX is a single 17MB binary that loads HuggingFace models directly and starts generating tokens.
No Python. No pip. No conda. No conversion tools.
Just one executable and your model weights.
Supported Models
- •Llama 3.x, 3.2
- •Qwen 2, 2.5, 3
- •Phi 3, 4
- •Gemma 2, 3
- •Mistral, Mistral Nemo
- •Granite (including MoE)
- •DeepSeek-R1 (distilled)
- •ModernBERT (embeddings)
What's Inside
- •Native BPE tokenizer with caching
- •16 chat template formats (auto-detected)
- •Quantized KV cache for long contexts
- •Metal-optimized inference via MLX-C
- •Streaming output with token callbacks
- •Zero-copy C interop from Zig
Why We Built It
We needed local inference that:
- Didn't require Python (dependency hell)
- Started instantly (no warm-up spinners)
- Compiled into our daemon (no subprocess overhead)
- Actually used Apple Silicon properly
Nothing fit. So we built VibeMLX in Zig, binding directly to Apple's MLX-C framework. 17MB. Zero dependencies. 32-66x faster time to first token than Python-based alternatives.
Context Engine
Samsara
Context That Never Dies
Named for the cycle of death and rebirth, Samsara ensures your conversations with AI never truly end. It maintains semantic continuity across sessions, intelligently compressing and retrieving context so your AI assistant always knows what you've discussed.
No more “context window exceeded.”
No more re-explaining your architecture.
No more starting over.
Conversation
• Current message
• Recent context
• Project files
Samsara Engine
• Semantic compression
• Vector search
• Intelligent retrieval
Retrieved Context
• Relevant history
• Project patterns
• Architectural context
Infinite History
Every conversation is indexed and searchable. Ask about something you discussed last week—Samsara remembers.
Project Awareness
File changes, dependency updates, structural patterns—all indexed and available for context retrieval.
Intelligent Compression
Old context is semantically compressed, not deleted. The meaning survives even when the tokens don't.
Seamless Retrieval
You don't manage context. Samsara does. The right information surfaces at the right time, automatically.
Templating Engine
vibe-jinja
Jinja2, Without the Jinja
Chat templates are critical for model compatibility. Every model expects a specific format—ChatML, Llama, Mistral, Phi, and dozens more. The standard solution is Jinja2, which means Python, which means dependencies.
We rewrote Jinja2 templating in Zig. Complete compatibility. Native performance. Zero external dependencies.
Because sometimes the right solution is the unhinged one.
What It Does
- •Full Jinja2 template syntax
- •Variable interpolation, filters, conditionals
- •Loop constructs with proper scoping
- •Template inheritance and includes
- •All the chat template formats models expect
Why It Exists
When your goal is a single binary with zero dependencies, “just pip install jinja2” isn't an option. vibe-jinja lets VibeMLX format prompts correctly for any model without touching Python.
Compatibility: Tested against the full Jinja2 test suite. If Python Jinja2 renders it, vibe-jinja renders it identically.
Terminal
Ghostty Inside
Yes, That Ghostty
Mitchell Hashimoto's GPU-accelerated terminal emulator—the one that broke the internet when it launched—is built directly into Vibe Mode. Not as a subprocess. Not wrapped in an iframe. Native integration via the libghostty embedding API.
You get the fastest terminal emulator available, and the whole application is still just 100MB.
Why It Matters
- •GPU-accelerated rendering — buttery smooth at any size
- •Native integration — shares memory with the app
- •Proper font rendering — your ligatures work correctly
- •Fast — really fast
- •It's Ghostty — you know it's good