Coze Brings AI Agents to the Desktop

🔑 Enhanced Key Takeaways

•Coze 2.0, launched on January 19, 2026, transformed the platform from a simple conversational AI tool into a comprehensive intelligent work partner capable of long-term planning, deep contextual understanding, and professional task execution.
•Coze 3.0, fully launched on June 1, 2026, introduced multi-person multi-agent collaboration, enabling team members to work with multiple agents simultaneously on complex tasks, with native support for OpenClaw access.
•The core code of Coze Studio and Coze Loop was officially open-sourced under Apache 2.0 on July 26, 2025, allowing for self-hosting and private deployment of the agent development and operations platform.
•Coze operates on a freemium model, offering a free tier with limited credits and paid plans (Premium Lite, Premium, Premium Plus) starting from $9/month for increased usage and advanced features.
•The deep screenshot feature, triggered by a hotkey, utilizes the operating system's Accessibility API to extract both the DOM structure tree and pixel image of the interface, converting visual information into precise UI descriptive text for language models.

📊 Competitor Analysis▸ Show

Feature/Pricing/Benchmarks	Coze	Claude Code	Manus My Computer	Perplexity Computer	ChatGPT Agent
Core Function	AI Agent development & deployment, desktop integration, multi-agent workflows, cloud drive, deep screenshot.	AI coding agent, desktop app, command center for autonomous coding agents.	Integrated Productivity & Content Creation, hybrid cloud-to-local.	Complex Research & Analysis, multi-model orchestration.	General Purpose Web Tasks, integrated with ChatGPT ecosystem.
Pricing Model	Freemium (Free, Premium Lite $9/mo, Premium $19/mo, Premium Plus $39/mo)	Paid (Pro $20/mo, Max from $100/mo)	Freemium (with paid tiers)	Paid (Part of Perplexity Pro)	Paid (Requires ChatGPT Plus/Pro)
Key Differentiators	No-code agent builder, Skills Marketplace, Agent Plan, multi-platform deployment, two-way cloud sync, deep screenshot for UI interaction.	Terminal-first, desktop-first approach, sandboxing for autonomous coding agents.	Hybrid cloud-to-local model with a focus on security.	Multi-model orchestration for deep research.	Seamless integration with the ChatGPT ecosystem.
Supported Models	GLM 5.3, Seedance 2.5, MiniMax H3, GPT, Claude, Gemini, Doubao, Tongyi, Hunyuan, Zhipu GLM, Kimi, DeepSeek.	(Implied Claude models)	(Not specified)	(Not specified)	(Implied GPT models)

🛠️ Technical Deep Dive

Deep Screenshot Technology: The feature captures both the pixel image and the underlying DOM structure tree and interaction logic of the current software interface by calling the operating system's Accessibility API. This allows Coze to convert unparsable images into precise UI descriptive text, enabling pure language models without multimodal visual processing capabilities to interact visually.
GLM 5.3 Model: This is Z.ai's (Zhipu AI's) latest flagship coding and agentic model. It reuses the 743B-parameter Mixture-of-Experts base from GLM-5.2, with performance gains derived from scaled-up post-training on diverse and complex task environments. It features a 1M-token context window and a maximum output length of 128K tokens, always operating with reasoning enabled at adjustable effort levels (low, high, max).
Seedance 2.5 Model: ByteDance's next-generation AI video model generates single-shot videos up to 30 seconds long in 1080p resolution (with reports suggesting up to 4K). It accepts up to 50 multimodal reference inputs (images, videos, audio) for consistent character and scene generation, offers 20% better prompt accuracy than its predecessor, and supports region-level editing and native audio with lip-sync for over 20 languages.
MiniMax H3 Model: An open-weights, general-purpose multimodal video model from MiniMax. It processes text, images, video, and audio inputs in a unified context to generate coherent audiovisual content. It supports text-to-video, first/last-frame image-to-video, reference generation, and precise video editing, with output resolutions up to 2K and durations of 4-15 seconds. Its architecture is a 33.1B dense, single-stream Omni Transformer.
Coze Platform Architecture: The backend of Coze is built using Go, while the frontend utilizes React and TypeScript, following a microservices and Domain-Driven Design (DDD) architecture for scalability and secondary development.
Agent Cloud Drive: Provides a dedicated, shared data storage space for users, project works, and multiple agents, supporting automatic cloud synchronization after local file modifications. The free version offers 4GB of storage, with enterprise expansion up to 20TB.

🔮 Future ImplicationsAI analysis grounded in cited sources

Coze's desktop client will significantly democratize AI agent development and deployment for a broader user base.

By integrating deep screenshot capabilities, cloud-drive synchronization, and a no-code/low-code platform, Coze lowers the technical barrier for users to create and deploy AI agents that interact directly with their local environment and files.

The platform will drive a new era of multi-agent, cross-device productivity in professional and personal workflows.

The combination of two-way cloud-drive sync, deep screenshot for contextual understanding, and automatic recognition of local agents like Codex CLI and Claude Code enables seamless, complex multi-agent collaboration across various devices and applications.

ByteDance is strategically positioning Coze to become a foundational ecosystem for AI agent development, challenging existing frameworks.

Through open-sourcing core components (Coze Studio, Coze Loop), integrating with its extensive infrastructure, and continuously upgrading with advanced models and features, ByteDance aims to establish Coze as a standard for building and managing AI agents.

⏳ Timeline

2023-Q4

Coze (international version) originally launched.

2024-Q1

Coze (Chinese version, '扣子') launched in China.

2025-04-18

ByteDance initiated beta testing for Coze Space, an AI-powered productivity platform.

2025-07-26

Core code of Coze Studio and Coze Loop officially open-sourced under Apache 2.0.

2026-01-19

Coze 2.0 officially launched, transforming into an intelligent work partner with Agent Skills and Agent Plan.

2026-06-01

Coze 3.0 fully launched, introducing multi-person multi-agent collaboration and OpenClaw support.

2026-08-19

Coze desktop client launched with Agent Cloud Drive, deep screenshot, and multi-device collaboration.

Coze Brings AI Agents to the Desktop

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (36)

👉Related Updates