Cindy AI Agent Explained: One Agent, Many Harnesses

8 min read · Last updated: October 24, 2026

Cindy is an open-source AI agent client, licensed under Apache-2.0, that runs on your own machine and wraps existing coding-agent harnesses, starting with Claude Code and Codex, behind one app. You can switch models and harnesses mid-task while memory, skills and tools carry over, and you can bring your own subscription, API keys or local models.

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The repo

• makecindy/cindy on GitHub: README, license, repo layout and privacy notes (read October 24, 2026)

Key takeaways

• Cindy's client is free and open under Apache-2.0, with a desktop app (Electron) and a mobile app (Expo / React Native) in one pnpm monorepo.
• The first 2 supported harnesses are Claude Code and Codex. More are being added, and a native harness is in the works.
• Models and harnesses can switch mid-task while your workspace, memory, skills and tools stay continuous.
• There are 4 ways to bring a model: Cindy's hosted service, the Claude Code / Codex Coding Plan you already pay for, your own API keys, or local models.
• "Skip Sign-In" runs local agents with no Cindy account, but server-backed features are unavailable.
• Building from source needs Node.js 22.x, pnpm 10.x (v11 is not yet supported) and Git LFS.

What is the Cindy AI agent, in one comparison?

Think of the Cindy AI agent as a general contractor. Claude Code and Codex are specialist crews, the models are the workers on each crew, and Cindy keeps one shared site notebook that every crew reads. You can swap crews halfway through a job, or run two at once, without losing the notes.

That notebook is the whole idea. The project's tagline is "Consider it done" (想到，就能做到 in the Chinese README).

A harness is the program that wraps a model and gives it a loop, file access and tools. Claude Code is one. Codex is another. Today, if you use both, each lives in its own house. What you taught one does not follow you to the other.

Cindy sits above them. Per the README, she "brings multiple harnesses, models and tools into one agent," and memory is "shared across harnesses." Correct her once and she does it right from then on, whichever crew is on site.

The contractor can also split a job. The README says one task "can even be planned, executed in parallel, and reviewed by agents on different harness × model combos." One crew plans, two build at the same time, a third inspects.

Contractor world	Cindy	What the README says
The contractor	Cindy, the client app	One agent that finishes real work in your projects and apps
Specialist crews	Harnesses	Claude Code and Codex first, a native harness in the works
Workers on each crew	Models	Mix freely with harnesses, switch mid-task
The site notebook	Memory, skills, tools, workspace	Stay continuous across switches
Standard procedures	Skills	Teach a way of working once, reuse it everywhere
Recurring jobs	Automation	Recurring work schedules itself, runs itself, reports back
Outside suppliers	MCP	Wire your internal tools and business systems in

If you read our explainer on OpenClaw, the personal AI agent, the difference is the layer. Cindy does not try to replace Claude Code or Codex. She orchestrates them.

How does Cindy work, step by step?

Cindy runs locally on your machine, using your real files and logged-in apps. You install the client, choose whether to sign in, pick where your model comes from, then hand her a task. She routes it to a harness and model, keeps memory and skills across switches, and can drive your browser, computer and phone.

Here is the flow, in order.

1. Install the client. Downloads are on cindy.app. Linux users get a separate guide for Ubuntu, Arch Linux and Omarchy that covers installation, updates, keyring setup and migration from a manual install.
2. Choose a mode. Sign in with a Cindy cloud account for the full hosted service, or pick "Skip Sign-In" on the login screen to run local agents. The app then shows "Not signed in."
3. Bring a model. Pick one of the four routes in the table below.
4. Give her a task. She works inside your projects, using your real files and the apps you are already logged into. She can also take work from IM (instant messaging) and from schedules.
5. Mix harness and model. They combine freely. Switch mid-task and the workspace, memory, skills and tools stay where they were.
6. Shape her over time. Correct her once and the memory sticks. Save a way of working as a skill. Add MCP (Model Context Protocol) servers for your internal systems. Install plugins through SkillHub or manually.

How do you pay for the model?

The client is free. The model is where cost lives, and the README lists four ways to bring one.

Route	What it means	Billing
Official Cindy service	Sign in to Cindy's hosted service	Usage deducted transparently, pricing on cindy.app
Your Coding Plan	Authorize the Claude Code / Codex Coding Plan you already pay for	"No duplicate bill"
Your own API keys	Connect keys you already hold	Billed by your provider
Local models	Run a model on your own hardware	No per-call bill

The local route pairs well with open-weight models. Our guide to running Google Gemma 4 locally is a good place to start if you have never done it.

What is the difference between hosted and Skip Sign-In?

Mode	Account requirement	What you get
Hosted service	Cindy cloud account	Cindy's full hosted service
Skip Sign-In	None	Local agents only, shown as "Not signed in"

One detail matters here. Skip Sign-In is not a connection to a local server. It just runs local agents without an account, and server-backed capabilities are off.

What is actually in the repo?

Path	What it holds
apps/desktop	Electron desktop client
apps/mobile	Expo / React Native mobile client
packages/*	Shared capabilities: auth, device-link, agent orchestration, model providers and more
apps/*-bin	Tool binaries for the desktop app. None are committed: claude-code, codex and ripgrep download per platform during pnpm install

The backend service is not here. It lives in a separate repository that is not part of this monorepo. By default the client talks to Cindy's official cloud, with endpoints listed in config/endpoint.json and config/endpoint.global.json, and desktop auto-updates come from the official CDN. The README says that is intentional, so outside developers do not need to self-host a server.

Does Cindy collect your data?

Official builds include TapDB usage analytics that record device, OS and app-version metadata, tied to your account ID after sign-in. Per the README, they do not collect chat content, file content or working-directory data. Signed-in clients also send a heartbeat with account ID, platform and version only.

Crash dumps stay on your machine and are never uploaded automatically.

If you build from source, you can drop analytics. Mobile is off by default: without TapDB credentials injected at build time, apps/mobile/src/analytics/mobileTapdb.ts does nothing. On desktop, remove the initTapdb() call in apps/desktop/src/renderer/index.tsx.

When should you use Cindy, and when should you not?

Use Cindy if you already pay for Claude Code or Codex and want one place where memory and skills follow you across both. Hold off, for now, if you need a fully self-hosted stack, harnesses beyond those two today, or features the README still lists as "in the making," like team skill sharing.

Use it if	Hold off if
You switch between Claude Code and Codex and are tired of re-teaching each one	You need a self-hosted backend: the server code is in a separate repo, not this one
You want to keep your existing Coding Plan with no second bill	You only use one harness and are happy with it
You want local models, or no account at all, via Skip Sign-In	You need server-backed features without signing in
You want an agent that can also drive your browser, computer and phone	You need to share skills with a team today: that is "in the making"
You want to audit or fork the client under Apache-2.0	You need a plugin marketplace now: that is "in the making" too
You are on Node.js 22 and pnpm 10	Your toolchain is locked to pnpm 11, which is not yet supported

My take: the value is not any single feature. It is that a correction you make on Monday in one harness is still there on Friday in another. If you never switch tools, that value is close to zero.

A small example: one task, three agents

Here is how a plan, parallel build and review could look, drawn in words. The pairings are my illustration. The README says these combos are possible, not which ones Cindy picks.

Task: "Add rate limiting to the /login route"

[1. Plan]      Claude Code × model A   → writes the plan
                        │
            ┌───────────┴───────────┐
[2. Execute]  Claude Code × model B    Codex × model C     (in parallel)
            └───────────┬───────────┘
[3. Review]    Codex × model A         → reviews the diff

Shared the whole time: workspace · memory · skills · tools

The point of the drawing is the bottom line. Every box reads and writes the same notebook, so the reviewer knows what the planner decided.

If you want to try it from source, this is the README's minimal entry point:

git clone https://github.com/makecindy/cindy.git
cd cindy
git lfs pull
pnpm install

Then start the desktop app against your own Cindy account. Pick the region on purpose; the README says not to rely on the internal default:

# Mainland China Cindy account
pnpm restart:desktop:remote --region=cn

# Global Cindy account
pnpm restart:desktop:remote --region=global

Remote development uses your existing login state, so you can continue sessions while you work on the client. If you send a pull request, every commit needs a Developer Certificate of Origin sign-off (git commit -s). A DCO check enforces it, and no CLA is required.

FAQ

Is Cindy free to use?

The client is free and its source is open under Apache-2.0. You pay for the model, not the app: through Cindy's hosted service with usage deducted transparently, through the Claude Code or Codex Coding Plan you already have, through your own API keys, or by running a local model.

Can I use Cindy without an account?

Yes. Choose "Skip Sign-In" on the login screen and Cindy runs local agents, showing your account state as "Not signed in." Server-backed capabilities are unavailable in that mode, and it does not connect you to a local server.

Which harnesses does Cindy support?

The first supported harnesses are Claude Code and Codex. The README says more are being added and a native harness is in the works. Models and harnesses can be mixed and switched mid-task.

Is the whole Cindy stack open source?

No. The desktop and mobile client plus shared packages are open under Apache-2.0. The backend service lives in a separate repository outside this monorepo, and model weights, datasets, prompts and trademarks may carry their own license terms.

Clone the repo, try Skip Sign-In with a local model first, and see whether shared memory earns its place in your week. Then browse the free tools on Art of Code, or read which coding skills survive when AI does the typing.

Written by Adithya Guttha, Founder of Art of Code.