Cindy is an Apache-2.0 agent client that runs Claude Code, Codex and Pi inside one desktop and mobile app, and it was trending on GitHub on October 24 with roughly 471 stars gained in a day and about 3.8k in total. According to the makecindy/cindy README, it "runs locally on your own machine, using your real files and logged-in apps," and it layers shared memory, skills, scheduling and MCP tools over the agents it wraps. Its tagline is "Consider it done." This article covers what it is, how to install it, what it costs, and what to check before you hand it your real files.
If you use more than one coding agent, the pitch will sound familiar. We have covered the same problem from several angles: cross-harness agent memory, shared sessions and handoff between agents, and a local control plane for running agents side by side. Cindy's difference is that it packages the answer as a product with a mobile app and a hosted option.
TL;DR: the questions people ask first
| Question | Answer |
|---|---|
| What is it? | A client that drives Claude Code, Codex and Pi, with shared memory, skills, schedules and MCP |
| License | Apache-2.0 for the client; model weights, prompts and trademarks may carry other terms |
| Is the backend open? | No. The README says it is in a separate, non-public repository |
| Cost | Free tier (bring your own key or subscription); Plus from $20 a month with a managed model service, per the project site |
| Platforms | Desktop on macOS, Windows and Linux; mobile on Android and iOS, where iOS is for viewing and taking over desktop sessions, per the site FAQ |
| Telemetry | Official builds include TapDB analytics; signed-in clients send a heartbeat |
| Sandbox? | README does not describe one; the site claims isolated workspaces and approvals |
| Maturity | Native harness, plugin system and marketplace are marked as in development |
What does Cindy actually do?
Cindy is not a new model and not a new coding agent. The README describes it as combining "multiple agent harnesses, models, and tools into one agent." The first supported harnesses are Claude Code and Codex, the site adds Pi, and a native harness is described as "in the works." On install, pnpm install downloads the tool binaries for your platform, namely claude-code, codex and ripgrep.
The features the README lists are:
- Memory shared across harnesses, so a correction you give once persists when you change agents.
- Reusable skills, the same pattern covered in our agent skills guide.
- Scheduled automation and work intake from chat tools.
- MCP integration for internal tools. If MCP is new to you, start with our MCP explainer.
- Switching models and harnesses mid-task, with the workspace staying continuous.
- Browser, computer and phone control.
The project site adds multi-agent "crews" (one model planning, three parallel workers, a separate reviewer), 20-plus built-in tools, integrations with Slack, Feishu, GitHub, Jira, GitLab, Confluence, Google Workspace and SSH, and remote control from mobile. These are the vendor's descriptions. We have not run them.
Why would you want a layer above the agents?
The main reason is switching cost. Claude Code and Codex each keep their own memory files, session history and conventions, so moving a task from one to the other means re-explaining context. A study we covered on what a harness actually buys you in tokens is a useful reminder that harness choice changes behavior and cost, which is why people want to try more than one on the same task.
A second reason is reach. If you want to approve a change from your phone or kick off a task by mentioning an agent in Slack, you need something that sits outside a single terminal session. Our piece on the terminal era giving way to the agent as the primitive argues the same shift is already underway.
The cost of a wrapper is one more dependency between you and the tools you already trust. If Cindy lags behind a Claude Code release, or changes how it launches the agent, you inherit that lag.
How do I install it?
The README gives two paths. For prebuilt apps, use the downloads on cindy.app; there is a separate guide for Linux covering Ubuntu, Arch and Omarchy. To build from source:
git clone https://github.com/makecindy/cindy
cd cindy
git lfs pull
pnpm install
You need Node.js 22.x, pnpm 10.x (the README says v11 is not yet supported) and Git LFS. The repository is a pnpm monorepo with an Electron desktop client in apps/desktop and an Expo and React Native mobile client in apps/mobile. Contributions need a DCO sign-off.
On first launch you can sign in to the hosted Cindy service, bring your own API keys, point it at local models, or authorize an existing Claude Code or Codex plan. "Skip Sign-In" runs local agents without an account but disables server-backed capabilities.
What does it cost?
The software is free. The project site lists these tiers:
| Tier | Price | What you get |
|---|---|---|
| Free | $0 | Software only; bring your own key or existing subscription |
| Plus | From $20 a month (monthly plans save 20%, per the site) | Managed model service with a monthly credit equal to the plan price; higher tiers exist |
| Team | Coming soon, waitlist | SSO, model access and credit controls |
| Enterprise | Contact sales | Self-hosted gateway in your VPC, audit logs, dedicated support |
If you already pay for Claude or Codex, the free tier is the relevant one, and Cindy adds orchestration rather than a second bill. The managed tier makes sense only if you want one invoice and no key management. We did not compare the credit value with direct API pricing.
What are the risks?
Because the question is "should I give this my real files," here is the honest list, all from the project's own README and site.
- Real-machine access. The README says it works with your actual files and logged-in apps. It does not describe a permissions or sandbox model. The site claims isolated workspaces, line-by-line review, rollback checkpoints and approval before irreversible actions, but that is marketing copy until you test it.
- Telemetry. Official builds include TapDB analytics. The README says it does "not" collect chat content, file content or working-directory data. Signed-in clients also send an online heartbeat with account ID, platform and version. Crash dumps stay local.
- Default cloud. Clients connect to Cindy's official cloud by default, and the backend is not public, so you cannot audit what it does. Self-hosting is described only for the Enterprise tier.
- Inference still leaves your machine. Files stay local, but requests go to whichever model gateway you configure, as with any hosted model.
- Credentials. The README warns "Never commit credentials or authorization files to the working tree." Authorizing an existing Claude Code or Codex plan means Cindy handles those tokens, so treat it like any app with access to your subscriptions. The recent Claude Code email-in-header leak is a reminder that even first-party clients leak identifiers in unexpected places.
- Unfinished parts. The native harness, plugin system, public marketplace and proactive suggestions are labeled as in development. The repository shows around 1.6k open issues and 309 pull requests, which signals heavy interest, and also a large backlog.
- Provider terms. Driving Claude Code or Codex through a third-party client with your own subscription may raise terms-of-service questions. Read the current terms for the plan you use before relying on it.
If you let any agent act on your machine, put a guard between the agent and dangerous actions. AgentBeam, built by the explainx.ai team, is an agent security platform that stops AI agents before they take dangerous actions.
How does it compare with other agent harness projects?
| Project | Approach | Our coverage |
|---|---|---|
| Cindy | Product-style client over Claude Code, Codex and Pi; hosted option; mobile | This article |
| Agent Relay | Shared sessions and handoff flows between agents, Apache licensed | Agent Relay |
| AgentPlane | Local control plane, side-by-side runs | AgentPlane |
| Beacon | Cross-harness memory only | Beacon |
| ECC | Large agent and skill bundle | ECC harness guide |
Cindy is the broadest of these and the most product-shaped. The others are narrower and easier to audit. If your only pain is lost context between agents, a memory layer alone may be enough.
A sensible way to try it
- Install from a release or build from source in a fresh machine user or VM.
- Choose "Skip Sign-In" with a local or BYO-key setup first, so no account data is involved.
- Point it at a throwaway repository, not your main one.
- Check network traffic once to see what the app contacts at idle.
- Test the checkpoint and rollback claim: make an edit, roll back, confirm the files return.
- Only then connect Slack, GitHub or other integrations, one at a time, with scoped tokens.
What this means for what you build or pay
Cindy is a signal that the market is moving from picking one coding agent to managing several. If you already run both Claude Code and Codex, a free client that carries memory and skills across them could save re-explaining time. If you do not switch agents, you gain little and add a closed-source backend dependency. The 471 stars in a day show interest; they do not show security, and nothing here is independently tested yet. We will update this article if the project publishes a sandbox description or if the backend is opened.
Related reading
- Agent Relay: shared sessions and flows for coding agents
- AgentPlane: a local control plane for Claude Code, Codex and Cursor
- Beacon: cross-harness agent memory
- What does a harness buy? Tokens, mostly
- The terminal era is over: the agent is the new primitive
- ECC agent harness install risks
- Official: makecindy/cindy on GitHub, cindy.app
Details are accurate as of October 24, 2026, based on the project README and website; prices and features can change.