AI AI Toolkit
Taking Off TodayTypeScriptAGPL-3.0

Let Web ChatGPT Safely Touch Your Local Files and Terminal: This Local Control Plane Ships 332 Tools and a Ledger

⭐ 483 Stars

What This Thing Actually Does

In one sentence, it gives web-based assistants like ChatGPT and DeepSeek a pair of local hands, letting them inspect, edit, build, and debug files and projects on your machine without exposing your whole filesystem or shipping proprietary code to untrusted cloud intermediaries. Nexus calls itself an enterprise-grade local AI control plane, and it works by bridging remote LLMs to a workstation through a Secure MCP Tunnel built on Streamable HTTP with Bearer tokens. The architecture has three layers. At the top sit ChatGPT, DeepSeek, and modern web assistants. In the middle is a Tauri 2 desktop control center running a Fastify MCP and REST server on port 18080 with SQLite WAL persistence. At the bottom is a local runner daemon that talks to the control center over bidirectional JSON-RPC 2.0 on WebSocket. All the real work happens in that daemon.

Why It Blew Up Recently

I think it hits a sharply growing tension. People want AI to write code, edit files, and drive software for them, yet they do not dare hand over their local projects wholesale. Connecting a web assistant straight to local files is risky, while a cloud sandbox cannot touch your real working environment. Nexus answers with a bridge that has boundaries: sandbox isolation, tiered authorization, command risk checks, and a ledger that survives crashes. That combination of wanting both capability and control is especially strong in enterprises, and with MCP becoming the de facto standard for tool access, this kind of local control plane naturally draws attention. The repo was created on November 13 and was closing in on five hundred stars within three days, helped by a clearly drawn architecture diagram, screenshots of the desktop control center, and a headline number of over 190 test suites, all of which give the impression that this was built seriously.

The Interesting Bits

The hardest part is the system of 332 canonical MCP tools. Every tool is formally defined in @localbridge/protocol with strict input schemas, risk ratings, and automated evidence tracking, spanning 24 capability dimensions: filesystem, advanced file operations, Git, command execution, persistent background jobs, LSP code intelligence, managed worktrees, browser automation, native Windows computer use through WinRT OCR and UI Automation, vision analysis, document engineering, local intelligence storage for memory, rules, knowledge and prompts, agent orchestration, and system discovery. The author calls this a zero-hallucination guarantee, meaning the tool definitions are the single source of truth and the model cannot invent a tool on the fly. The Action Ledger WAL is another highlight. Before every critical tool execution, a write-ahead log is recorded, so after a crash or reboot the system can recover pending operations, restore a consistent workspace state, and use an idempotency guard to avoid duplicate execution during flaky reconnects. Security is handled through multi-tier project and sandbox isolation: granular read and write permissions, path isolation, AST-level command risk controls, and access modes that switch between read-only and read-write plus execution modes that switch among disabled, safe-only, and project-code, all without restarting. Credentials are encrypted with DPAPI, and it manages MCP processes such as Figma, GitHub, SQLite, and Filesystem. On top of that, it bridges generative AI into Blender 4.x and 5.x for procedural modeling and PBR shaders, and even handles Unity scenes.

Who It Is For

Heavy users of AI coding assistants who worry about code leakage should try this first, since it lets a web ChatGPT operate local projects safely. Developers who put privacy first and insist on local deployment will like the sandbox plus ledger design. Windows users get extra treatment: native computer use, Job Objects, and WinRT OCR are Windows-specific strengths. Enterprise security teams can study its authorization model and audit ledger, and even if they never deploy it, the architecture is worth reference. One warning: permission setup has a learning curve. You authorize a project in the desktop control center, set access and execution modes, then generate an MCP Client Token and paste it into your AI client, and that flow takes a little time the first round.

Quick Start

The easy path is the release page, where you download the installer for your operating system: the x64 setup.exe for Windows, universal or Apple Silicon dmg for macOS, deb or AppImage for Linux, plus a Tampermonkey userscript for browser users. After install, Nexus lives in your system tray with a bundled Node.js runtime and background daemon. To build from source, prepare Node.js 24 or newer, pnpm 10 or newer, and stable Rust, then git clone, pnpm install, pnpm run build, and finally pnpm --filter @localbridge/desktop dev. Connecting an AI client takes three steps: authorize a project under Workspace then Projects and set its modes, generate an MCP Client Token under Settings then Tokens, and paste http://127.0.0.1:18080/mcp with the Bearer token into a client such as a ChatGPT custom GPT action.

How It Compares

  • Cloud AI coding tools that run in a hosted sandbox: convenient, but they cannot touch your real local environment and your code has to be uploaded.
  • Mounting a local folder directly to a web AI: maximum capability but almost no boundaries, with high risk of mistakes and leaks.
  • Generic MCP clients: they connect tools but lack a unified tool definition, risk ratings, sandboxes, and a crash-recovery ledger, so enterprise-grade governance is mostly absent.
  • Nexus: 332 tools with a single source of truth, plus sandboxes, tiered authorization, DPAPI-encrypted credentials, and a WAL ledger, offering an engineered balance between capability and control.
  • The catch: AGPL-3.0 carries network-distribution obligations; the token and permission flow is a bit complex for newcomers; the scale of 332 tools means real documentation reading before you are productive; and Windows-specific strengths are weaker on macOS and Linux.
🚀

Get Started

Open Source · Commercial Friendly

AGPL-3.0· TypeScript