• Node.js or Python running the application logic.
• External vector databases (Chroma, Pinecone, Qdrant) running as separate network services.
• Background LLM daemons (Ollama, vLLM) communicating across localhost ports.
• The localhost IPC tax — every hop between app, vector database, and daemon serializes JSON over a socket.

We built Sofuu to collapse this entire architecture into one in-process runtime, a single ≤ 5.5 MiB binary.

What is Sofuu?

Sofuu is a general-purpose runtime for AI — built to run anywhere from servers to smart devices. It is not an npm framework or a wrapper around Node: it is a standalone native binary with a Rust/C systems core (QuickJS + libuv, zero JIT) that measures a 22.5 ms median cold start (wall clock including process spawn, Apple M2 Pro — every number in this post is measured and published in the repo's BENCHMARKS.md).

TypeScript and JavaScript are the orchestration surface. The primitives live in the runtime core:

1. In-Process Cognitive Vector Memory (QTSQ)

Built-in encrypted .qtsq memory containers with hardware SIMD search — measured 0.82 ms for top-K over 10,000 × 768-dim vectors on Apple Silicon, in-process, with no external database. And QTSQ itself is now open source under the Apache-2.0 license (github.com/haruhito2666/qtsq).

2. Autonomous Agent Loops & Prompt Sovereignty

Define multi-step agents with native tools, budget caps, cancel signals, and streaming observability. Sofuu guarantees prompt sovereignty: the runtime never injects hidden system prompts or biases.

3. Native Model Context Protocol (MCP)

First-class MCP client and server support is built into the standard library. Connect to existing tool ecosystems without installing adapter packages.

4. Sandbox v2 & 5-Shield Security

Untrusted agent tools and third-party plugins cross five shields: schema validation, policy and network jailing, a circuit breaker, watchdog deadlines, and a panic guard — a throwing plugin becomes a typed error, never a dead process.

Honesty note: the built-in inference engine is experimental. For production workloads, point the built-in llamacpp provider at a llama.cpp server.

Hello World in Sofuu

Memory, an agent, and recall — one process, no node_modules:

// Memory + agent + tools in one process
const brain = sofuu.memory.open("./brain.qtsq", 768);
brain.remember(
  sofuu.ai.embedLocal("Deployment target: AWS us-east-1"),
  "Deployment target: AWS us-east-1"
);

sofuu.agent.define({
  name: "deploy-auditor",
  system: "Inspect the repository and summarize deployment requirements.",
  tools: ["web", "memory"],
  budget: { maxSteps: 12 },
});

const result = await sofuu.agent.run(
  "deploy-auditor",
  "Audit this repo for deployment requirements"
);
console.log(result.answer);

const [hit] = brain.recall(sofuu.ai.embedLocal("deployment target"), 1);
console.log("Recalled:", hit.text);

Run it immediately:

sofuu run agent.ts

(No compilation step — TypeScript is stripped in-process. No node_modules. No background daemons.)

Runs Everywhere, Ships as an SDK

The same binary runs on macOS and Linux (including Raspberry Pi 5), and the engine embeds into iOS and Android apps through the libsofuu C ABI — QuickJS has no JIT, so App Store policy holds by construction. The Android SDK publishes as a Gradle dependency:

implementation 'com.sofuu:sofuu-android:0.3.0'

Open Source — All of It

The Sofuu runtime is MIT-licensed, and the QTSQ memory engine is open source under the Apache License 2.0.

• Website & Documentation: https://sofuu.xyz
• GitHub Repository: https://github.com/sofuu-runtime/sofuu
• QTSQ Engine: https://github.com/haruhito2666/qtsq
• Official AI Skill: https://sofuu.xyz/SKILL.md

Try it in 10 seconds:

# macOS & Linux (auto-detects architecture) — or: npm i -g sofuu
curl -fsSL https://sofuu.xyz/install | sh

Current release: v0.3.0-beta.2. We are actively developing Sofuu and would love your feedback, bug reports, and benchmark tests!