makecindy/cindy

The problem

If you already use Claude Code or Codex, you know the friction: each harness has its own workspace, its own context, and its own idea of what you taught it last week. Work that spans your files, browser and phone means stitching together separate sessions by hand. Cindy packages those harnesses into a single local agent that runs on your own machine with your real files and logged-in apps, so context and tooling carry across tasks instead of resetting at every switch.

What's different

The usual answer is to pick one CLI agent and script around it, or to assemble a framework like LangGraph and build the orchestration yourself. Cindy's bet is orchestration at the client layer: harnesses and models are interchangeable, switchable mid-task, and one job can be planned, executed in parallel and reviewed by agents on different harness × model combinations. Memory, skills and MCP connections are shared across those harnesses rather than owned by one. It also ships as an actual Electron desktop app plus an Expo/React Native mobile client, which is a different posture from terminal-only agents.

Who it's for

Developers and technical operators who want an agent on their own machine that touches real files and apps, and who value being able to swap harnesses or bring their own keys. It suits people already paying for a Claude Code or Codex plan, since the README claims that plan carries over. Look elsewhere if you want a headless, embeddable agent runtime for a server product — this is a client, and the backend is explicitly not in this repository — or if you need a thin library rather than a full app.

Cost of adoption

Node.js 22.x, pnpm 10.x (v11 unsupported) and Git LFS are required; tool binaries for Claude Code, Codex and ripgrep are downloaded per platform during install. The client defaults to Cindy's official cloud endpoints and auto-update CDN, so a self-hosted deployment is not a supported path. "Skip Sign-In" gives local agents only, with server-backed features unavailable. Official builds include TapDB analytics (device/OS/version metadata, tied to account ID) and a signed-in heartbeat; the README documents how to strip that when building from source. The README is detailed about architecture and build steps but says little about the maturity of the plugin marketplace or team skill sharing — both are described as "in the making."

Verdict

Cindy is a young project: created roughly three months ago, at about 4,000 stars, with a single day on the trending board (+471 on 2026-10-24). That is a spike pattern, not an established track record, and the absence of any Hacker News discussion means independent scrutiny is thin. The multi-harness orchestration idea is genuinely more interesting than another chat wrapper, and Apache-2.0 with DCO-only contribution is a friendly setup. But the value proposition leans on the hosted service, and the repo is a client whose server half you cannot see or run. Treat it as worth a trial install if you already live in Claude Code or Codex; treat it as unproven infrastructure for anything you need to depend on.