ByteBuster-NeuroSyncOS：本地优先的隐私优先 AI 项目编排器

NeuroSync Sovereign OS

A local-first, privacy-first AI project orchestrator. It runs as a single Node.js server with a React dashboard, schedules background agents, and routes requests across local and free-tier LLM providers -- all state lives in a local SQLite database, not a third-party cloud service.

What's here

The system is organized into eight modules, each with its own dashboard:

PortGrid -- capability/tool governance, the workflow approval queue, and an embedded project terminal
ScopeLogic -- guided interview engine that turns a request into a DAG proposal
RouteSwitch -- LLM provider registry (API keys persisted in SQLite), fallback chains, free-tier-aware routing, and a Council Mode arbitration/confidence log
BaseVault -- the SQLite data layer, backup/restore, redaction settings, and project storage with background sync
CoreExec -- the DAG workflow engine, scheduler, crash recovery, and cron jobs
ScoutDaemon -- idle-time background research and hardware monitoring
Cerebro -- conversational memory with decaying vector search and a learning-approvals (human-in-the-loop) pipeline
CEPH -- a unified observability layer (logs, metrics, traces, health) that consolidates the dashboard and the API behind one abstraction

Backend: Node.js 22, Hono, better-sqlite3 (WAL mode), sqlite-vec for local embedding search, and workers via poolifier. Frontend: React + Vite.

Prerequisites

Node.js 22 or newer
npm

Getting started

npm install

# Terminal 1: API server (http://localhost:3743)
npm run dev:server

# Terminal 2: UI dev server (http://localhost:3742)
npm run dev:ui

The server boots a FreeLLMAPI-compatible provider setup by default: an OpenAI-compatible base URL and API key live in the RouteSwitch provider registry and persist in the local SQLite database, so the UI's RouteSwitch Set-up screen can configure and test providers without needing env vars.

RouteSwitch also supports a router-side ensemble: request model="fusion" on a chat completion and the router fans the request out and returns one answer. NeuroSync does not run that ensemble -- it passes the model id through and, for OpenAI-compatible and FreeLLMAPI providers, reports the X-Routed-Via response header so the dashboard can show which upstream served the call. Optional per-request fields can be supplied via extraBody (a trusted-caller setting from config or a provider row); it can never override the model, the prompt, or the structured output format, and prototype-key names (__proto__, constructor, prototype) are refused outright. See docs/architecture/FUSION.md for the full contract, the authoritative denied-key list, the provider-by-provider coverage, and the time budgets involved.

The UI isbuilt to use the 2026 Unified MCP standard. When deployed as a Tauri bundle it ships its own 32-bit sidecar wrapper; the remainder of the stack -- server, UI dev server, tests -- all run plain under Node.js on a desktop or an edge host.

Testing

npm test              # run the test suite (vitest)
npx tsc --noEmit       # typecheck
npm run build          # production build

Documentation

Further documentation lives under docs/, including:

User manual
Security & privacy model, threat model
Architecture blueprint
Contributor templates (ADRs, RFCs, review checklists)

See SECURITY.md for how to report a vulnerability, and CONTRIBUTING.md for how to set up a dev environment and submit changes.

License

Apache License 2.0