The most interesting new agent project on GitHub this week does not build an agent. It builds a shell around other people's agents.
Cindy, an Apache-2.0 desktop and mobile client, picked up roughly 1,300 stars in a day according to GitHub Trending. The repository contains no agent core: the first supported harnesses are Claude Code and Codex, with a native harness described in the README as "in the works". Models and harnesses can be mixed freely and switched mid-task while the workspace, memory, skills and tools stay continuous. A single task can even be planned, executed in parallel and reviewed by agents running different harness-model combinations.
What is in the box
According to the README, Cindy runs locally on your machine, against real files and already signed-in apps, and can drive the browser, the computer and a phone; work can be kicked off from messaging apps and schedules. The desktop app is Electron, the mobile app is Expo/React Native, and the monorepo is pnpm with Node 22. The client source is open; the backend is not. It lives in a separate repository, and official builds connect to Cindy's cloud services by default.
You pay either through the hosted service, with usage deducted, or through a coding subscription you already own. The project says you can authorise an existing Claude Code or Codex plan inside Cindy and keep using it "no duplicate bill"; your own API keys and local models also work. A "Skip Sign-In" mode runs local agents without a Cindy account, at the cost of the server-backed features.
Why this is the story
Cindy is not an outlier, it is the visible edge of a shift. JetBrains shipped ThinkRail, an IDE that is essentially a thin host around the `pi` agent, running against the user's own provider credentials. Another trending repository, ECC, stacks skills, memory and guardrails on top of Claude Code, Codex and OpenCode. TradingAgents, a multi-agent framework for financial analysis, added sign-in with your own ChatGPT Plus or Pro plan in version 0.7.0.
Innovation is moving up a layer. The model is not the moat and neither is the agent loop; the workspace, the memory, the permissions and the question of whose subscription pays the bill are.
The awkward questions
Three of them matter. First, terms of service: using one vendor's subscription inside another vendor's product is precisely the grey zone where providers like to change the rules retrospectively, and none of these projects has a public agreement with the plan owners. Second, security: an agent allowed to drive a browser, a computer and a phone, and to load plugins from a marketplace, is a far larger attack surface than a chat window. Third, observability: open client, closed backend — and TapDB telemetry on by default in official builds, which the project says carries no chat, file or working-directory content.
To its credit, Cindy documents all of that plainly and can be run without an account, which is more than many comparable clients manage. Anyone trying it should still start with a scratch directory and no production credentials.
The caveat
A project with a few days of public history and one day of trending is not an established tool, and a README describes intentions, not audit results. But the direction is hard to miss: the next round of agent tooling will not compete on which model answers. It will compete on who owns the session the answer arrives in.
