The fragmentation of AI coding agents just got a unified front. Cindy, a new open-source project hosted on GitHub, introduces a harness that runs Claude Code and Codex as interchangeable engines. Instead of juggling separate CLIs and subscriptions, developers get a single interface to switch between Anthropic and OpenAI’s agentic workflows without losing session context.

Swappable Harnesses and Multi-Device Sync

Cindy’s core architecture treats model providers as swappable components. The documentation explicitly states that switching harnesses "hands off or rebuilds the engine context" rather than forcing internal state identity. This allows users to start a session on a work computer, continue it on a phone via remote desktop, and even leverage model providers configured on a different machine. The system supports a third harness, Pi, a lean open-source agent, alongside the major commercial ones.

Local-First with Optional Cloud Services

Unlike heavy server-dependent tools, Cindy operates on a local-first model where the user's computer acts as the server. This eliminates the need for server rental or complex self-hosting setups for the core functionality. However, the architecture is hybrid: cloud companions can sync conversations, memory, and skills through Cindy’s services, while model requests are still processed by the chosen provider. Project files remain on the local device, ensuring remote access does not equate to whole-disk synchronization.

Developer Experience and Extensibility

Built as a pnpm monorepo with Electron for desktop and Expo/React Native for mobile, Cindy emphasizes extensibility through a SkillHub and plugin system. The project is licensed under Apache-2.0, encouraging community contributions and forks. Developers can define recurring routines, connect tools via MCP, and create custom companions with specific roles. The client is free to use with existing subscriptions, though managed model services and provider billing still apply.

Key Takeaways

  • Cindy unifies Claude Code, Codex, and Pi into a single interface with swappable backends.
  • The architecture is local-first, using the user's computer as the server to avoid rental costs.
  • Sessions can be resumed across devices (macOS, Windows, Linux, iOS, Android) without file copying.
  • The project is open-source (Apache-2.0) and extensible via a SkillHub and plugin ecosystem.

The Bottom Line

This is a pragmatic solution for the 'subscription fatigue' plaguing AI developers. By treating Claude and Codex as interchangeable engines, Cindy lowers the barrier to switching providers mid-workflow, though the reliance on separate cloud sync services for full multi-device continuity is a necessary trade-off for the local-first promise.