H4D3ZS/vscodium-rust
A local-first, agentic IDE built on Rust/Tauri v2 backend and React 19/TypeScript frontend, with a VS Code–shaped editor but native process management instead of Electron Local inference via Lemonade (an OpenAI-compatible llama.cpp server on port 13305), with optional two-model split using a large ROCmFPX reasoner and a small qwen3.5:4b operator Kortex submodule (AGPL-3.0) provides KV-slot cache, AIM retrieval proxy, and VFS daemon for context optimization and semantic search over the workspace
Analysis
TL;DR
- A local-first, agentic IDE built on Rust/Tauri v2 backend and React 19/TypeScript frontend, with a VS Code–shaped editor but native process management instead of Electron
- Local inference via Lemonade (an OpenAI-compatible llama.cpp server on port 13305), with optional two-model split using a large ROCmFPX reasoner and a small qwen3.5:4b operator
- Kortex submodule (AGPL-3.0) provides KV-slot cache, AIM retrieval proxy, and VFS daemon for context optimization and semantic search over the workspace
- Speculative decoding support via prompt-lookup n-grams and MTP heads, with live acceptance rate monitoring and prefill-savings measurement tools
- Full iOS build pipeline on Windows/Linux without macOS/Xcode, plus APEX specialist routing, secret scanning, and PyTorch/ONNX export capabilities
Why It Matters
This project represents a significant shift toward local-first AI development environments, demonstrating that agentic IDE capabilities can run entirely on consumer hardware (including AMD GPUs) without cloud dependency. The architecture choices—native Rust backend, speculative decoding, and retrieval-augmented context routing—offer a blueprint for privacy-conscious AI tooling that reduces latency and API costs while maintaining full model control.
Technical Details
- Architecture: Rust/Tauri v2 backend with React 19/TypeScript frontend; editor loop, indexing, and process management run in native processes rather than Electron's main thread
- Inference Stack: Lemonade serves as the sole local backend (OpenAI-compatible, port 13305) running llama.cpp directly with ROCmFPX for AMD hardware; optional two-model split separates a large reasoner from a fast operator (qwen3.5:4b default)
- Kortex Services: KV-slot cache (:1537) implements KDKVC to skip re-prefilling repeated prompt prefixes; AIM retrieval proxy (:1536) builds dense workspace catalogs and gates context chunks by relevance; VFS daemon (:1538) manages .aimmemory and file watching
- Agentic Features: Multi-turn tool loop with verify-before-done semantics, shadow workspace for safe edits, background agents for long-running work, inline agent-edit diffs with per-hunk accept/reject
- Mobile & Security: Cross-platform iOS build (ARM64 Mach-O, .ipa packaging, zsign/ldid signing) from Windows/Linux; APEX specialist routing, headless-browser automation, secret scanning, PyTorch train loop with ONNX export
Industry Insight
- The local-first agentic IDE model challenges the cloud-dependent trajectory of AI coding tools, suggesting that on-device inference with optimization layers (KV caching, speculative decoding, retrieval gating) can achieve competitive UX at lower marginal cost
- The two-model split architecture (large reasoner + small operator) offers a practical template for balancing capability and latency in agentic workflows, and the open KDKVC approach could influence how future IDEs handle context window management
- Cross-platform iOS toolchain without Xcode lowers the barrier for mobile development workflows in AI-assisted environments, potentially accelerating adoption of agentic patterns in mobile CI/CD pipelines
Disclaimer: The above content is generated by AI and is for reference only.