Cindy is an open-source AI agent client for desktop and mobile that runs coding harnesses such as Claude Code and Codex locally, against your own files. It is for developers who already keep two or three agent tools open and lose their notes, fixes and context every time they move between them. The repository, makecindy/cindy, is the client side of the project: the desktop and mobile apps plus their shared packages, organized as a pnpm monorepo. It is TypeScript, Apache-2.0 licensed, and has about 6,100 stars and 759 forks since its first push in July 2026.
What it does
Cindy presents itself as one agent that finishes real work inside your projects and apps, not a chat window you copy answers out of. According to the project's README, it:
- runs locally on your machine, using your real files and your logged-in apps
- brings multiple harnesses, models and tools together under a single agent
- supports Claude Code and Codex as its first harnesses, with more being added and a native harness described as in the works
- can drive your browser, your computer and your phone
- can take work from instant messaging and from schedules
The channel's summary adds the part that sells it: correct Cindy once on something and she does it that way from then on, and that memory follows you when you change harness. All of this is the project's account of its own software, not an independent measurement.
How it works
The interesting design choice is that Cindy treats the harness as swappable. Models and harnesses mix freely and, per the README, can switch in the middle of a task while your workspace, memory, skills and tools stay continuous. One task can be planned by one agent, executed in parallel, and reviewed by another, each on a different harness and model combination. That is why the project calls itself a client: the intelligence is whatever you connect, and Cindy is the shell that keeps state across those connections.
The repository's shape matches the platform list in its topics: Electron for desktop, React Native for the mobile apps, with shared TypeScript packages between them in a pnpm monorepo.
Getting started
The topics name macOS, Windows, iOS and Android as targets, so this is meant to be a real app on all four rather than a terminal tool. The project does not list hardware requirements, memory or GPU anywhere in this material, so treat the system needs as undocumented.
Money is the part to be clear about. The client is free to use and its source is open under Apache-2.0, but you still pay for the intelligence: Cindy drives Claude Code and Codex, and those need your own paid plan or API key. The README begins a sentence about bringing models your way by signing in, but the saved copy cuts off at that point, so the exact list of sign-in and key options cannot be quoted here.
To get it: prebuilt apps are linked from the project's own download page at cindy.app/download/, and the source is at github.com/makecindy/cindy. Building it yourself means a monorepo checkout with the toolchain its badges name, Node 22.x and pnpm 10. No single install command appears in the material, so do not expect a one-liner.
When to use it / when not
The main catch is scope: this repository is the client only, and the project is young, with its first push in July 2026 and commits continuing through October. Anything beyond the client, the native harness included, is described as planned or in progress rather than shipped, and you are putting a three-month-old codebase between you and tools you already depend on.
Use it instead of opening Claude Code or Codex directly in a terminal when you regularly switch between them and want one memory, one set of skills and one mobile surface across both. Don't use it if your work lives inside a single harness you are happy with, or if you need a stable, boring tool, since this adds a moving part on top of the agent you already trust.
Alternatives
The alternative most readers already have is the harness itself: Claude Code or Codex, run from a terminal or inside an editor. That adds no new layer to debug, and it is what Cindy sits on top of. What it does not give you is memory shared across harnesses, parallel subtasks reviewed by a second agent, or a phone client. If none of those three matter to you, you do not need Cindy.
This is worth a look if you pay for more than one coding agent and keep re-teaching each of them the same corrections. The pitch is narrow and concrete: keep the memory, swap the engine. The star count shows there is an audience, but the maturity is what you would expect from a project this new, so try it on work you can afford to supervise before you let it drive your browser and your phone.