# Sofuu > Sofuu is a self-contained runtime for AI workloads. One native binary bundles QuickJS (ES2023, no JIT), a libuv event loop, WHATWG streams, in-process GGUF/SafeTensors inference on Metal/NEON/AVX2 (experimental engine), autonomous agent loops, a native Model Context Protocol client and server, and encrypted on-device vector memory (`.qtsq`, with opt-in Qdrant/Pinecone/pgvector adapters). Selected local workflows need no separate Python runtime or daemon. OS libraries, model files and chosen providers or servers may be needed. The macOS arm64 binary is 5.39 MiB; the Linux builds (5.88 MiB arm64, 6.80 MiB x86_64) currently exceed the project's 5.5 MiB release cap, which blocks a green release. Kaze is a domain-specific language for bounded agent and device routines. Its current embedded host can return synthetic AI, tool and memory results; a real model-and-authorized-tool journey is not yet established. Plugins run in-process and must be trusted as code. - Version: 0.3.0-beta (active development; APIs may change before a stable release) - License: MIT core; the QTSQ memory engine is open source under the Apache License 2.0 (https://github.com/haruhito2666/qtsq) - Source: https://github.com/sofuu-runtime/sofuu - Website: https://sofuu.xyz - Official binaries: macOS arm64, Linux x86_64, Linux arm64. Windows x86_64 is not CI-covered and the MSVC build of vendored QuickJS is fix-forward. - Install: `curl -fsSL https://sofuu.xyz/install | sh` or `npm install -g sofuu` ## What it is - A general-purpose JavaScript/TypeScript execution environment: `sofuu run app.ts`, `sofuu eval`, an interactive REPL, an HTTP server, a bundler, and an npm installer. - Cold start — process invocation to a completed `1+1` — measures 22.5 ms median wall clock including process spawn (31 runs) on an Apple M2 Pro / macOS 15.7.3 with the current tree (BENCHMARKS.md, 2026-10-31); the published 0.3.0-beta artifact measures 61.5 ms the same way. `sofuu version`, which loads no engine, is a 5.0 ms process floor. The smaller figures still printed in BENCHMARKS.md predate the current engine and that file marks them historical. - An in-process AI substrate: local LLM inference (experimental built-in engine; for production point the `llamacpp` provider at a llama.cpp server), embeddings, speech, vector math, agents, tools, MCP, and memory are runtime primitives, not packages. - An embeddable engine: the same core ships as `libsofuu` with C, Swift, and Kotlin bindings for iOS, Android, and macOS apps. QuickJS has no JIT, so iOS App Store executable-memory rules hold by construction. Build the C ABI from source today (`make libsofuu`); the SwiftPM and CocoaPods coordinates do not resolve at `0.3.0-beta` and land with the next release. ## Who it is for JavaScript and TypeScript developers building offline-first or private AI features, CLI tools, local agents, MCP servers, edge services, robotics, and embedded/mobile apps who want one process and no daemon stack. ## When to use it - Local or offline LLM inference inside a JS/TS app (experimental engine; production: a llama.cpp server through the `llamacpp` provider) without Ollama, Python, Docker, or CUDA. - Agents and MCP servers using supported policy, approval, budget, and cancellation controls; enforcement varies by execution path. - Local vector memory is encrypted by default, with explicit plaintext mode and unencrypted brain-card exports. `embedLocal` uses a 768-dimensional trigram hash rather than a trained semantic model; remote adapters are opt-in. - Single-binary distribution with no runtime installation, no telemetry, and no background processes. - Embedding an AI-capable JS engine into iOS/Android/macOS/C apps. ## When not to use it - Training, fine-tuning, or multi-GPU distributed inference — use PyTorch/CUDA or a serving framework. - Maximum single-model tokens per second — llama.cpp is ahead of Sofuu on both prefill and decode in the published parity ledger (https://github.com/sofuu-runtime/sofuu/blob/main/contract/SCOREBOARD.md). - Applications that require full Node.js API parity or native V8 N-API addons such as `sharp`, `canvas`, or `better-sqlite3`. - Production Windows deployments today; the MSVC build of vendored QuickJS is fix-forward. ## Install - [Install script](https://sofuu.xyz/install): `curl -fsSL https://sofuu.xyz/install | sh`, verifies SHA-256 and installs to a user-writable path. - [Downloads](https://sofuu.xyz/downloads): macOS arm64, Linux x86_64, Linux arm64 tarballs with checksums, LICENSE, and NOTICE. - [npm](https://www.npmjs.com/package/sofuu): `npm install -g sofuu` or `npx sofuu`. ## Guides - [AI applications](https://sofuu.xyz/ai): local inference, embeddings, agents, memory, and the 10-second hello world. - [Runtime](https://sofuu.xyz/runtime): engine architecture, cold start, CLI, sandboxing, and the embeddable C ABI. - [MCP](https://sofuu.xyz/mcp): native Model Context Protocol client and server, transports, config, and policy. - [Compare](https://sofuu.xyz/compare): honest trade-offs against Node.js, Bun, Python + Ollama, llama.cpp, and the LangChain/Chroma stack. ## Documentation - [Docs](https://sofuu.xyz/docs): installation, CLI, JavaScript API, MCP, RLM, ML gates, benchmarks, and embedding. - [README](https://github.com/sofuu-runtime/sofuu#readme): the canonical feature and API reference. - [Conformance](https://github.com/sofuu-runtime/sofuu/blob/main/contract/CONFORMANCE.md): the 52-clause contract board per platform, live capability posture, and stability index. - [Scoreboard](https://github.com/sofuu-runtime/sofuu/blob/main/contract/SCOREBOARD.md): the llama.cpp inference-parity ledger, updated from real sweep receipts. ## Machine-readable - [SKILL.md](https://sofuu.xyz/SKILL.md): developer skill for AI coding assistants that read SKILL.md (for example Claude, Cursor, Copilot, Antigravity — nominative, no affiliation implied). - [llms.txt](https://sofuu.xyz/llms.txt): this file. - [llms-full.txt](https://sofuu.xyz/llms-full.txt): complete API signatures, CLI reference, runnable recipes, and the capability matrix. - [sitemap.xml](https://sofuu.xyz/sitemap.xml) - [robots.txt](https://sofuu.xyz/robots.txt) ## Project - [GitHub](https://github.com/sofuu-runtime/sofuu) - [QTSQ memory engine](https://github.com/haruhito2666/qtsq) — open source, Apache-2.0 - Contact: hello@sofuu.xyz