Cindy

• Using an existing Claude Code or Codex plan for general desktop work such as organizing files, editing documents and spreadsheets, or working through signed-in web apps, without paying for a second AI subscription
• Having Claude Code plan a code change, several Codex or Claude Code workers implement and test it in parallel, and a different model review the diff
• Scheduling recurring jobs such as daily issue triage, weekly reports or competitor monitoring, with results sent to Slack, Feishu or another IM
• Starting a task from your phone or a Slack thread, then approving risky steps or entering a verification code from the phone while it runs on your computer
• Running the same agent setup on local models through Ollama or vLLM when data should not reach an external model provider
• Connecting internal tools through MCP so the agent can work with company systems alongside files and browser sessions

Cindy's bet is that the harnesses people already trust for coding, Claude Code and Codex, are good enough to run general computer work if someone wraps them in a proper desktop app. That wrapper is substantial: harness and model switching mid-task, line-by-line review and checkpoints, schedules, MCP, Slack and Feishu delegation, and mobile apps that can approve steps or hand back a login. Its handling of Claude subscriptions is careful and documented, which matters for anyone who wants to reuse a plan they already pay for. The caveats are about maturity and openness. The client is Apache-2.0, but the backend is not, the app talks to Cindy's cloud by default, and official builds carry analytics. The product is at 0.1.x and changes almost daily, much of its documentation is in Chinese, and team features are not out. It is a strong fit for an individual developer or power user who wants one agent across coding and everyday computer work on their own keys or plan; organizations wanting a fully self-hosted, stable stack should wait.

About Cindy

Cindy is a desktop agent app from XD Inc. (the iOS and Android listings are published by XD Entertainment Pte Ltd) whose open-source client repository was created on GitHub in July 2026. Rather than building its own agent loop first, it wraps existing agent harnesses: Claude Code and Codex ship with the desktop app, Pi is supported as a third harness, and a Cindy-native harness is described as in progress. You pick a harness and a model per task and can switch either mid-task while the workspace, memory, skills and tools carry over; a collaboration mode can have one harness and model plan a task, several workers execute in parallel, and another combination review the result. Work happens on your machine, in your projects, files, terminal and signed-in browser sessions, with browser, computer and Android control, SSH to remote servers, and connectors for Slack, Feishu, GitHub, Jira, GitLab, Confluence and Google Workspace. The guardrails are those of a coding tool applied to general work: changes start in an isolated workspace, every edit goes through a line-by-line review panel, checkpoints let you roll back, cost is shown as a task runs, and the agent is meant to stop and ask before irreversible actions. Around that sit persistent memory shared across harnesses, reusable skills, scheduled automations that report back to chat, custom MCP tools, and companions: long-lived named agents with their own memory, group chats and IM bots that can import a personality, memories, skills and automations from Hermes or OpenClaw. The iOS, iPadOS and Android apps are remote controls for tasks on your computer: dispatch work, watch progress, approve sensitive actions or take over through remote desktop. The business model separates software from inference. The client is free and every software capability works with your own API key, a local model (OpenAI-compatible endpoints, Ollama, vLLM) or a Claude Code or Codex plan you already pay for. For a Claude plan, Cindy runs the bundled Claude Code CLI's own sign-in and leaves the credentials in that CLI's store; its docs cite Anthropic's rule that subscriptions may only be used through unmodified Claude Code. The optional paid plans (PLUS $20, SUPER $100, ULTRA $200 a month, 20% less on an auto-renewing plan) buy matching usage credit on Cindy's managed model gateway. The tradeoffs: only the client is open source (the backend lives in a separate repository and the app connects to Cindy's cloud by default), official builds include TapDB usage analytics, the product is at version 0.1.x with near-daily releases, release notes and much of the developer documentation are in Chinese, and the Team plan is not available yet.