Cindy：跨 Harness 持久记忆的本地优先开源 AI Agent，纠正一次、处处生效

Tech-Experiment

Jul 27, 2026

🇨🇳 中文

“Correct her once; she remembers across all harnesses.”

这句话说的是：在 Claude Code 里纠正了一个习惯，下次切到 Codex，同样生效。

这不是一个小改进。当前 AI Agent 生态里，每个 Harness 都是孤岛——Claude Code 有自己的 CLAUDE.md，Codex 有自己的配置，它们之间没有共享的记忆层。每次切换，都要从零开始。

Cindy 想解决的就是这个问题。

一、它是什么

Cindy 是一个本地优先的开源 AI Agent 客户端，Electron 桌面端 + Expo/React Native 移动端，TypeScript 写成，Apache-2.0 授权。

2026 年 7 月 22 日上线，5 天 776 star。

核心定位：Harness 无关的 Agent 编排层。把 Claude Code、Codex 这些已有的 Harness 接进来，在它们之上建立统一的记忆、Skills 和多 Agent 协作能力。

不是”套壳”——它没有把这些 Harness 替换掉，而是作为一个协调层，让它们能够共享状态。

二、跨 Harness 持久记忆

这是 Cindy 最核心的设计。

现在每个 Harness 都有自己的记忆层：Claude Code 读 CLAUDE.md/skills，Codex 读自己的配置，互不相通。用户纠正 Claude Code 的行为，切到 Codex 又要重新说一遍。

Cindy 的做法：在所有 Harness 上层建立统一的记忆存储。Memory 在这里不是某个 Harness 的配置文件，而是 Cindy 客户端层面的持久状态，跨 Harness 共享：

用户纠正一次
    ↓
Cindy 记忆层
    ├── 注入到 Claude Code session
    └── 注入到 Codex session

“Correct her once” — 纠正动作只发生一次，对所有 Harness 生效。

除了 Memory，还有 Skills（一次定义、处处复用的工作方式模板）。技术上是同一个层的两种表达：一种存矫正记录，一种存可调用的工作流程。

三、Orca 多 Agent 编排

AGENTS.md 里有一条规则：

修改 Orca 多 Agent 协同时，必须先读 docs/dev-rules/orca-team-architecture.md

README 的描述：

“one task can even be planned, executed in parallel, and reviewed by agents on different harness × model combos”

一个任务，不同阶段可以分配给不同 Harness × Model 的组合：

Plan   → Claude Opus 4.8 (Cindy native harness)
Execute → Claude Code × Sonnet (并行执行)
Review → Codex × GPT-5.5 (独立评审)

各个 Agent 的 Workspace、Memory、Skills 和 Tools 保持连续——切换 Harness 不等于切换上下文，Memory 层保证了连续性。

四、技术架构

仓库结构

apps/
  desktop/     → Electron 桌面客户端
  mobile/      → Expo / React Native 移动端
packages/
  *            → auth, device-link, agent orchestration, model providers...
cindy-protocol/  → wire 协议（git submodule，与服务端共享）

pnpm monorepo，Node.js 22.x。

Harness 的可执行文件（claude-code、codex、ripgrep）不提交进仓库，由 pnpm install 按平台下载。这意味着你本地已有的 Claude Code / Codex 安装可以直接被 Cindy 使用。

本地优先

Cindy 跑在你自己的机器上，访问你真实的文件系统和已登录的 Apps。

数据不上传（除了可以在源码里关掉的 TapDB 聚合统计）。

隐私设计：聊天内容、文件内容、工作目录数据不被收集。Crash dump 留在本地不自动上传。

服务端与客户端的分层

Cindy 的开源部分是客户端。服务端在独立仓库，未开源。

使用方式有三种：

Cindy 官方云服务（托管 API，按用量计费）
直接用你已有的 Claude Code / Codex 订阅（不产生双重计费）
本地模式（无需 Cindy 账号，不连接服务端）

五、接入已有的 Claude Code / Codex 订阅

这一点值得单独说。

如果你已经订阅了 Claude Code 或 Codex，可以在 Cindy 里直接授权它们，不产生额外费用——Cindy 调用的是你自己的 Coding Plan 配额，不是另起一套计费。

git clone https://github.com/makecindy/cindy.git
cd cindy
git submodule update --init --recursive cindy-protocol
pnpm install
pnpm restart:desktop:remote --region=global  # 用 Cindy 云账号

本地模式只需要在登录界面选 “Local mode”，不创建 Cindy 账号。

六、技术判断

Cindy 切入的问题是真实存在的：

当你同时用多个 Harness 时，维护跨 Harness 一致性的成本是真实的。 CLAUDE.md、Codex 配置、各种 skills 文件需要分别维护，纠正一个 Harness 的习惯不会自动迁移到另一个。

Cindy 的解法——在 Harness 之上建统一的记忆层——在逻辑上是对的。它不替换 Harness，而是作为协调层存在，这也是为什么它可以在不改变你已有订阅的情况下接入。

需要观察的是：

服务端闭源 意味着某些能力（device-link、IM 集成、自动化调度）依赖 Cindy 的服务端。对于完全自主可控的用户，这是个依赖项。本地模式可以绕开这个依赖，但功能也相应受限。

Orca 多 Agent 编排 目前只在文档里提到架构，实际产品化程度还需要实际使用验证。

总体判断：对于每天在多个 Harness 之间切换的用户，Cindy 解决了一个真实的痛点。值得装上试一试——毕竟 Apache-2.0，风险可控。

项目：github.com/makecindy/cindy（776 ⭐，Apache-2.0）
网站：cindy.app
下载：cindy.app/download
语言：TypeScript，pnpm monorepo
平台：macOS/Windows（Electron）+ iOS/Android（React Native）

🇬🇧 English

Cindy: Local-First Open-Source AI Agent with Cross-Harness Persistent Memory

“Correct her once; she remembers across all harnesses.”

What this means: correct a habit in Claude Code, and it applies the next time you switch to Codex.

This isn’t a small improvement. In today’s AI agent ecosystem, every harness is an island — Claude Code has its own CLAUDE.md, Codex has its own configuration, with no shared memory layer between them. Every switch means starting over.

Cindy aims to solve this.

What It Is

Cindy is a local-first open-source AI agent client — Electron desktop + Expo/React Native mobile, built in TypeScript, Apache-2.0 licensed.

Launched July 22, 2026, 776 stars in 5 days.

Core positioning: harness-agnostic agent orchestration layer. It brings Claude Code, Codex, and other harnesses into a single client, building unified memory, Skills, and multi-agent collaboration on top of them.

Not a “wrapper” — it doesn’t replace these harnesses, but acts as a coordination layer so they can share state.

Cross-Harness Persistent Memory

This is Cindy’s core design.

Currently, every harness has its own memory layer: Claude Code reads CLAUDE.md/skills, Codex reads its own config, with no cross-talk. Correct Claude Code’s behavior, then switch to Codex, and you have to explain it again.

Cindy’s approach: build a unified memory store above all harnesses. Memory here isn’t a config file for some harness — it’s a persistent state at the Cindy client layer, shared across harnesses:

User corrects once
    ↓
Cindy memory layer
    ├── Injected into Claude Code session
    └── Injected into Codex session

“Correct her once” — the correction happens once, takes effect everywhere.

Beyond Memory, Skills (reusable workflow templates defined once and called anywhere) live in the same layer: one expression stores correction records, the other stores callable workflows.

Orca Multi-Agent Orchestration

AGENTS.md includes this rule:

Before modifying Orca multi-agent coordination, read docs/dev-rules/orca-team-architecture.md

README’s description:

“one task can even be planned, executed in parallel, and reviewed by agents on different harness × model combos”

One task, different phases assigned to different harness × model combinations:

Plan   → Claude Opus 4.8 (Cindy native harness)
Execute → Claude Code × Sonnet (parallel execution)
Review → Codex × GPT-5.5 (independent review)

Each agent’s workspace, memory, skills, and tools remain continuous — switching harnesses doesn’t mean switching context, because the memory layer guarantees continuity.

Technical Architecture

Repository Structure

apps/
  desktop/     → Electron desktop client
  mobile/      → Expo / React Native mobile
packages/
  *            → auth, device-link, agent orchestration, model providers...
cindy-protocol/  → wire protocol (git submodule, shared with server)

pnpm monorepo, Node.js 22.x.

Harness executables (claude-code, codex, ripgrep) are not committed to the repo — they’re downloaded per-platform by pnpm install. This means your existing Claude Code / Codex installations can be used directly by Cindy.

Local-First

Cindy runs on your own machine, accessing your real filesystem and logged-in apps.

Data doesn’t leave your device (except for the TapDB aggregate analytics, which can be removed in source builds).

Privacy design: chat content, file content, and working directory data are never collected. Crash dumps stay local and are never automatically uploaded.

Client vs. Server Split

The open-source part of Cindy is the client. The server lives in a separate repository and is not open-sourced.

Three modes of use:

Cindy official cloud service (hosted API, usage-based billing)
Use your existing Claude Code / Codex subscription (no duplicate billing)
Local mode (no Cindy account required, no server connection)

Using Your Existing Claude Code / Codex Subscription

This point deserves its own mention.

If you already subscribe to Claude Code or Codex, you can authorize them directly in Cindy without additional charges — Cindy draws from your own Coding Plan quota, not a separate billing line.

git clone https://github.com/makecindy/cindy.git
cd cindy
git submodule update --init --recursive cindy-protocol
pnpm install
pnpm restart:desktop:remote --region=global

Local mode just requires selecting “Local mode” on the login screen — no Cindy account needed.

Technical Verdict

Cindy addresses a real problem:

When you use multiple harnesses simultaneously, the cost of maintaining cross-harness consistency is real. CLAUDE.md, Codex config, and various skill files need to be maintained separately; correcting one harness’s behavior doesn’t automatically migrate to another.

Cindy’s solution — building a unified memory layer above harnesses — is logically sound. It doesn’t replace harnesses; it acts as a coordination layer. This is also why it can integrate without changing your existing subscriptions.

Points to watch:

The server is closed-source, meaning some capabilities (device-link, IM integration, automation scheduling) depend on Cindy’s server. For users who want full control, this is a dependency. Local mode bypasses this dependency but with reduced functionality.

Orca multi-agent orchestration is currently only described in documentation; the actual level of productization needs real-world validation.

Overall: for users who switch between multiple harnesses daily, Cindy solves a real pain point. Worth installing and trying — Apache-2.0 means the risk is manageable.

Project: github.com/makecindy/cindy (776 ⭐, Apache-2.0)
Website: cindy.app
Download: cindy.app/download
Language: TypeScript, pnpm monorepo
Platforms: macOS/Windows (Electron) + iOS/Android (React Native)

💬 评论与讨论

使用 GitHub 账号登录后发表评论