Nexus links web AI assistants to local files over MCP
A new TypeScript project called Nexus aims to let web AI assistants work on local files and commands through a secured MCP bridge, with an action ledger and sandbox controls.
A local bridge for web AI
As of 2026-11-17, the Nexus repository shows 362 stars and 64 forks. The project was created on 2026-11-13, so it gathered those numbers in four days. Nexus is mainly TypeScript and uses the AGPL-3.0 license. It describes itself as a local AI control plane for ChatGPT, DeepSeek and modern web AI assistants, with a companion browser extension called webagent named in the announcement. The pitch behind the Nexus local AI control plane is simple: keep the model remote, but give it a controlled route to a developer's own files and tools.
How the MCP tunnel works
The README says remote LLMs reach a local workstation through a Secure MCP Tunnel built on Streamable HTTP with Bearer token authentication. The desktop control center uses Tauri 2. A Fastify MCP and REST server listens on port 18080 and stores data in SQLite with WAL mode. A local runner daemon talks to it over bidirectional JSON-RPC 2.0 WebSocket.
The runner exposes 332 canonical MCP tools across 24 capability areas. Those areas include filesystem access, advanced filesystem operations, Git, command execution, persistent background jobs, TypeScript and Python language server intelligence, managed worktrees, browser automation, Windows computer use, vision analysis, document engineering, local memory and knowledge, and agent orchestration. The README says the tools are defined in a protocol package with strict input schemas and risk ratings.
Execution is meant to be recoverable and bounded. An action ledger writes critical tool calls to a write-ahead log before side effects for crash recovery and idempotent replay. A command risk engine classifies commands as SAFE, CAUTION or DANGEROUS and blocks destructive calls such as rm -rf /, format and reg delete by default. Windows Job Objects enforce process isolation, CPU quotas and memory caps. A node-pty terminal daemon keeps stateful shell sessions.
Security controls include path containment inside authorized project roots, blocking of symlinks and Windows NTFS junctions that point outside those roots, and separate 256-bit tokens for MCP clients and the runner daemon. The README says tokens are stored as SHA-256 hashes and compared with crypto.timingSafeEqual. A sensitive file shield masks .env files, keys, certificates, .git and cloud credentials.
Setup and controls
The source build requires Node.js 24 or newer, pnpm 10 or newer and Rust stable. The README gives commands to clone the repository, install workspace dependencies, build and launch the desktop app. It also mentions a desktop installer option when a verified build is available, but warns readers not to assume a Windows, macOS, Linux or userscript asset exists unless it appears on the v1.3.1 release page.
To connect an AI client, the user authorizes a project in the desktop control center, sets access to read-only or read-write, and sets execution to disabled, safe-only or project-code. The README says these can be switched without restarting. The user then generates an MCP client token beginning with lb_ and configures the client to use http://127.0.0.1:18080/mcp with an Authorization Bearer header. The README shows examples for ChatGPT and Custom GPT Actions, Cursor, Windsurf and any MCP client, using MCP-Protocol-Version 2024-11-05.
Limits and open questions
The source is a repository README, not an independent security review. Claims such as crash recovery, isolation and the tool definitions are presented by the project. The announcement names webagent but the supplied README text does not explain how the browser extension works or which browsers it supports. Platform support is also unclear. Windows computer use and Windows Job Objects are prominent, and the README explicitly says not to assume installers for other systems exist. Anyone enabling read-write, project-code, computer use or unrestricted mode takes responsibility for commands and file changes, according to the security notice. The AGPL-3.0 license also has network source obligations that teams should review before offering a modified version as a service.
What it means for building apps with AI
Nexus is a local control plane for AI clients that speak MCP, not a hosted app builder. Its useful lesson for builders is architectural: keep execution near the files, put a token and a path boundary in front of it, classify risky commands and keep an audit trail. For app builders, Draftship is a separate AI app builder that writes, builds and hosts apps with a built-in database, user sign-in and hosting. Paid plans add private projects, a choice of AI model, source code access, custom domains and GitHub push. Its docs explain how the AI chat and model selection work.
Primary sources
- ysunyang4/Nexus: Web AI for local work:让网页版 AI 直接操作你的本地文件和电脑,配套浏览器插件 webagent · GitHub trending AI repositories, Nov 17, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.