Model Context Protocol server for AI VRAM calculation and GPU recommendation

Accurate local-LLM memory math — will this model fit your GPU or Mac? Models sliding-window, hybrid/linear, MLA compressed-KV and MoE architectures that naive VRAM calculators get wrong by up to ~18x. Zero deps, single file, conformance-vector tested.

VRAM JSON cache with IndexedDB integration, built with TypeScript.

Enterprise-grade VRAM management for HTML5 games. Features hardware tier detection, priority-based hysteresis eviction, Safari throttle detection, and safe scene streaming.

Detect your hardware and find LLMs that will actually run on it — with real VRAM/KV-cache fit math, not opaque scores.

Can my machine run this model? Estimate VRAM/RAM, tok/s, and TTFT.

GPU VRAM probing — memory usage tracking, allocation monitoring, OOM prevention

CLI preflight checks for model VRAM, CUDA, GPU compatibility, and quantization risk.

Local low-resource voice extension for Pi featuring sub-6GB VRAM TTS/ASR with 4-bit quantization.

Run local models in opencode. llama.cpp as a provider, with live GPU, VRAM and CPU telemetry in the sidebar.

Run LLMs locally in your terminal. Supports custom .pt GPT-2 and LLaMA decoder checkpoints (auto-handles compact key naming, tied embeddings, and split-halves RoPE) plus GGUF models, with a live dashboard of throughput, VRAM, and chat.

PCSX-ReARMed libretro core (Sony PlayStation, HLE BIOS) for romdev, bundled as WASM with romdev's CPU + SPU + GPU-VRAM debug exports.

Otto Brain - self-contained host for local GGUF models, with measured VRAM budgeting and reasoning-budget control

Hardware-aware local-LLM loadout planner: chooses model class, quant, KV-cache, context, and offload policy for your VRAM/RAM, and refuses configs that would silently page to disk.

Zero-GC off-thread ImageBitmap loader with strict VRAM limits, URL deduplication, and LRU eviction to prevent mobile browser crashes.

MCP server that exposes a running unhosted daemon's local-first capabilities (memory, web fetch, vram-pool status) as tools to MCP-aware clients like Claude Desktop.

PineForge decentralized compute provider agent CLI

Zero-GC WebGL2 render-path telemetry. Deterministic command counters (draw calls, instances, floats uploaded) recorded in the GL-agnostic core, plus an async EXT_disjoint_timer_query_webgl2 GPU-time pool. Passive instrument; pairs with lite-vram (the mana

Native CLI worker for the opencompute inference network. Connects to the orchestrator over Socket.io and serves jobs from your GPU.

CLI interface to glTF Transform