Neo Chat：本地优先的 AI 聊天工作台，把「加密跨设备同步」这个硬骨头啃下来了

by Mycelium Protocol

---

项目地址：https://github.com/u14app/neo-chat 授权：MIT

---

一句话结论

Neo Chat 是一个把”本地优先”做到细节里的 AI 聊天工作台：聊天记录、工作区元数据、技能、插件配置、记忆、搜索索引、文件——默认全部留在浏览器里，服务端路由只是受控代理（转发模型请求、网页搜索、RAG 解析、语音、插件和 MCP 执行、部署健康检查）。Next.js 16 + React 19 + TypeScript 6 + Zustand，MIT 协议，1774 star。

它最值得写的不是”又一个开源聊天界面”，而是 v2.4.0 解决了本地优先应用里最难的那道题：多设备之间怎么同步，同时不破坏”数据不出本机”这个承诺。

加密跨设备同步：本地优先应用的死穴，它啃下来了

“本地优先”和”多设备可用”天然打架——数据留在本机，换个设备就没了；上云同步，“本地优先”就成了一句空话。

Neo Chat 的解法是一个可选的、端到端加密的个人保险库，通过 WebDAV 或 S3/MinIO 做跨设备同步。官方原话说得很直白：恢复密钥、凭据、本地基线数据、搜索缓存、向量嵌入——这些东西永远不会进入远程对象或 ZIP 文件。 也就是说，就算你用的 WebDAV/S3 服务商本身不可信，它能拿到的也只是加密后的密文容器，解密材料从来不会离开你的设备。

这是”local-first”和”跨设备”这对矛盾里，工程上最难受的一段——不是加个密就完事，而是要精确划清”哪些东西可以离开设备（加密后的数据块）“和”哪些东西绝对不能离开设备（密钥、凭据、明文索引）“这条线。

一套壳，接遍所有模型和工具

多供应商聊天：Google、Anthropic（原生 Messages API SDK 接入）、OpenAI、以及 OpenAI 兼容端点，供应商级别的作用域隔离。

远程 MCP + 本地 stdio 桥接：从官方 MCP Registry 直接发现和安装远程 streamable HTTP MCP 服务器，带插件市场管理、鉴权、服务端工具执行。本地 stdio 桥接跑在一个经过身份验证的 Docker 容器里，命令集合由部署配置固定死，不能运行时任意扩展——这是个明显有意为之的安全边界设计。

参数化 Skill：纯文本 Skill，支持带参数、支持把最多 4 个 Skill 打包成有序执行链，调用元数据可复现。

RAG：collection 级别的分块控制、Markdown 标题感知的预览、显式重建索引、混合词法/向量检索，向量检索不可用时优雅降级回词法检索而不是直接报错。

离线可用是真的离线

内置一个严格”零 API 缓存边界”的离线 PWA——不是那种”离线时显示缓存的旧结果”的伪离线，是明确划定哪些操作在没有网络时依然可用。配合本地全局搜索、多语言（英/中/日）设置搜索、首次运行模型引导，整个产品在断网状态下依然是一个可用的工作台，而不是一个报错页面。

谁该看这个

适合：想要一个功能完整的聊天工作台（多模型/RAG/MCP/记忆/语音全都要），但又不想把数据交给某个云服务商的人；需要多设备同步、但同步渠道本身信不过（公共 WebDAV、第三方 S3 兼容存储）的场景。

不适合 / 需要注意：本地 stdio MCP 桥接的命令集合是部署时锁死的，如果你需要运行时动态注册本地工具，这条路走不通，需要用远程 MCP 服务器代替。

---

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

TL;DR

Neo Chat is an AI chat workspace that takes “local-first” seriously down to the details: chat history, workspace metadata, skills, plugin configuration, memory, search indexes, and files all stay in the browser by default; server routes are controlled proxies only (forwarding model requests, web search, RAG parsing, voice, plugin/MCP execution, deployment health checks). Next.js 16, React 19, TypeScript 6, Zustand. MIT licensed, 1,774 stars.

What makes it worth writing about isn’t “another open-source chat UI” — it’s that v2.4.0 solved the hardest problem in local-first apps: how to sync across devices without breaking the “data never leaves your machine” promise.

Encrypted cross-device sync: the local-first Achilles’ heel, actually solved

“Local-first” and “multi-device” are naturally in tension — keep data on one machine and it’s gone when you switch devices; sync it to the cloud and “local-first” becomes an empty phrase.

Neo Chat’s solution is an opt-in, end-to-end encrypted personal vault synced across devices via WebDAV or S3/MinIO. The project states it plainly: recovery keys, credentials, local baseline data, search caches, and vector embeddings never enter remote objects or ZIP files. Even if the WebDAV/S3 provider you use isn’t trustworthy, all it ever gets is an encrypted container — the decryption material never leaves your device.

This is the genuinely hard engineering part of reconciling “local-first” with “multi-device” — not just bolting on encryption, but precisely drawing the line between “what may leave the device (encrypted blobs)” and “what must never leave the device (keys, credentials, plaintext indexes).”

One shell, every model and tool

Multi-provider chat: Google, Anthropic (native Messages API via the official SDK), OpenAI, and OpenAI-compatible endpoints, with provider-scoped isolation.

Remote MCP plus a local stdio bridge: discover and install remote streamable-HTTP MCP servers directly from the official MCP Registry, with plugin-market management, authentication, and server-side tool execution. The local stdio bridge runs in an authenticated Docker container whose command set is fixed by deployment configuration — not dynamically extensible at runtime. A deliberate security boundary.

Parameterized Skills: plain-text skills that accept parameters, chainable into ordered bundles of up to four, with reproducible invocation metadata.

RAG: collection-level chunking controls, Markdown heading-aware previews, explicit reindexing, hybrid lexical/vector retrieval that gracefully falls back to lexical search instead of erroring out when vector retrieval is unavailable.

Offline that’s actually offline

A built-in offline PWA with strict no-API-cache boundaries — not the fake kind of offline that just shows stale cached results, but one that clearly defines which operations actually work without a network. Paired with local global search, localized (English/Chinese/Japanese) settings search, and first-run model guidance, the whole workspace stays usable while disconnected, rather than degrading into an error page.

Who should look at this

Good fit: anyone who wants a full-featured chat workspace — multi-model, RAG, MCP, memory, voice, all of it — without handing their data to a cloud vendor; scenarios needing multi-device sync where the sync channel itself isn’t trusted (public WebDAV, third-party S3-compatible storage).

Not a fit / worth noting: the local stdio MCP bridge’s command set is fixed at deploy time — if you need to register local tools dynamically at runtime, this path doesn’t support it; use a remote MCP server instead.

---

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.