cindy

Product Introduction

1. Definition: Cindy is an open-source, multi-harness AI agent software designed for developers and technical teams. It functions as a unified orchestration layer that integrates various AI model providers (like Anthropic's Claude, Codex, Pi) and their specific agent frameworks (harnesses) into a single, continuous workspace.
2. Core Value Proposition: Cindy solves the critical problem of AI tool fragmentation and context loss. It enables users to leverage multiple best-in-class AI models and harnesses within a single task, switching between them seamlessly without restarting work, thereby increasing productivity and completion rates for real project work.

Main Features

1. Multi-Harness, Multi-Model Orchestration: Cindy's core architecture allows it to route tasks across different AI agent harnesses (Claude Code, Codex, Pi) and the models they support. This enables advanced workflows like using one model for planning, multiple models for parallel execution, and another for review, all within a single task context. The workspace, memory, skills, and tools remain persistent across switches.
2. Local-First, Isolated Workspace: Cindy operates with a local-first philosophy. Your project files, code, and chat history primarily reside on your computer. When performing edits, it uses an isolated workspace to prevent direct overwrites. All changes are presented in a review panel for line-by-line approval, and you can roll back to previous checkpoints, ensuring control and safety.
3. Extensible Tool & Skill System: The platform is built for extensibility. It comes with over 20 built-in tools for local workspace, terminal, collaboration (Slack, GitHub, Jira), and remote access (SSH). Users can extend capabilities via Model Context Protocol (MCP) to connect internal systems, create reusable custom skills, and will soon be able to use plugins to modify UI and behavior, fostering a customizable AI assistant.
4. Asynchronous & Cross-Platform Control: Cindy agents run persistently on your primary machine (desktop). You can monitor progress, ask follow-ups, and approve actions remotely from a mobile app (iOS/Android). You can also delegate tasks by mentioning Cindy (@cindy) in connected collaboration tools like Slack or Feishu; the task executes on your computer and the result is posted back to the chat thread.

Problems Solved

1. Pain Point: Developers and teams waste time context-switching between disparate AI coding assistants (like separate Claude Code, Cursor, and GitHub Copilot workflows), each with its own interface, memory, and limitations. This fragmentation hinders complex task completion and forces suboptimal model choices for entire tasks.
2. Target Audience: The primary users are software developers, engineering teams, DevOps engineers, and technical product managers who integrate AI into their daily development, debugging, documentation, and project management workflows. It's also suited for enterprises needing a governable, extensible AI agent platform.
3. Use Cases:
   • Parallelized Development Workflow: Architect a feature with Claude Code, implement it using Codex with Grok, write tests with Kimi, and have GPT review the diff—all in one orchestrated task.
   • Continuous Code Review & Refactoring: Delegate scheduled tasks to run linting, dependency updates, or security scans across codebases using different specialized models, with results reported to a team channel.
   • Unified Cross-Platform Assistance: Start a complex debugging session on a local machine, continue monitoring and providing guidance via mobile while away, and delegate a follow-up fix via a Slack command.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros:
     • Harness Agnosticism: Unprecedented flexibility to use the best model/harness for each subtask, breaking vendor lock-in.
     • Persistent Context: Maintains task memory and state across model switches, a significant advantage over using individual tools.
     • Developer Control & Safety: The local-first, review-required, and rollback-capable workflow builds essential trust for real work delegation.
     • Strong Extensibility: MCP and skill system allow deep integration into custom tech stacks.
   • Cons:
     • Complexity Overhead: The power requires understanding different harnesses and models; it's overkill for simple, single-model tasks.
     • Dependent on External Services: While it can use local models, optimal use relies on access to various proprietary API keys (Claude, GPT, etc.), managing costs and quotas.
     • Early-Stage Ecosystem: Some features like the native Cindy harness, public plugin marketplace, and proactive briefs are still "brewing" or "in development."

2. Key Alternatives & Differentiation:

   • Cursor / Windsurf: These are powerful AI-native IDEs but are primarily locked into their own agentic workflows (mostly using Claude or GPT). Cindy differentiates by being IDE-agnostic and allowing orchestration across the models these tools use individually.
   • Claude Code / Codex / Pi (Standalone): Using these harnesses directly means working in separate, siloed applications. Cindy's core value is unifying them, allowing a task to leverage Claude Code's planning, Codex's execution, and Pi's conversational review without manual copy-pasting.
   • OpenAI Assistants API / LangChain: These are lower-level frameworks for building custom agents. Cindy is a ready-to-use, opinionated product built on such principles. It offers a polished UI, built-in tools, and zero-config startup, whereas the frameworks require significant development effort.

Frequently Asked Questions (FAQ)

1. Is Cindy a replacement for Claude Code or GitHub Copilot? No, Cindy is an orchestrator that includes them. You authorize your existing Claude Code or Copilot (Codex) subscription within Cindy. It then allows you to use those harnesses alongside others, with shared memory and tools, rather than replacing them outright.
2. Where does my code and data go when using Cindy? By default, your project files and code remain on your local machine. Cindy's AI Gateway (which you can self-host) routes inference requests to the configured model APIs, but your workspace files are not uploaded to Cindy's cloud for storage, aligning with its local-first security model.
3. How does pricing work for Cindy AI agent? The Cindy client software is free and open-source (Apache 2.0). You can use it for free by bringing your own API keys. The paid "Plus" plan ($20/month) is for Cindy's managed AI Gateway service, which provides convenient, consolidated access to various model APIs (GPT, Claude, etc.) and includes a monthly usage credit.
4. Can enterprises self-host the entire Cindy platform? Yes, for enterprise needs. You can self-host the Cindy AI Gateway inside your own VPC for complete data control and compliance. The client software can then connect to this private gateway, allowing full audit logs, cost attribution, and use of internal models.
5. What is the difference between a "harness" and a "model" in Cindy? A harness (like Claude Code, Codex) is the specific agent framework or application logic that defines how a model works (e.g., its planning steps, tool use patterns). A model (like GPT-4, Claude 3.5 Sonnet) is the underlying AI. Cindy lets you mix and match, e.g., running the "Claude Code" harness with a "GPT-4" model, if supported.