Western observers of the AI agent ecosystem see a world dominated by LangChain, CrewAI, AutoGen, and a handful of coding agents. This view misses an entire parallel universe. China’s AI agent ecosystem — shaped by the Great Firewall, government regulation, and a distinct set of domestic model providers — has produced at least 15 significant open-source projects with combined GitHub stars exceeding 350,000.

Several of these dwarf their Western counterparts. Dify alone has 150K stars — more than LangChain and CrewAI combined. MetaGPT’s 69K stars would make it one of the most-starred agent projects in the world. And almost nobody in the Western agent community talks about them.

This is the story of how that parallel universe came to exist, what it looks like, and why it matters.

The Wall That Built an Ecosystem

The Great Firewall does more than block websites. It fundamentally shapes how Chinese developers build AI agents.

In the West, a developer building an agent framework starts with a simple assumption: OpenAI’s API is the default backend. This assumption pervades everything — LangChain’s deep OpenAI integration, CrewAI’s model-agnostic but OpenAI-first design, the entire tool-calling standard built around OpenAI’s function calling format.

Chinese developers cannot make this assumption. Direct access to OpenAI and Anthropic APIs requires VPN connections that are legally grey, technically unreliable, and impossible for enterprise deployments. The result is a domestic model ecosystem that has matured into a genuine alternative: GLM (Zhipu AI), Qwen (Alibaba), DeepSeek, Kimi (Moonshot AI), ERNIE (Baidu), and Doubao (ByteDance). Each ships its own SDK, its own function calling format, and its own fine-tuned variants for agent workloads.

This isn’t a limitation. It’s a selection pressure. Chinese agent frameworks must support multiple model providers from day one — a multi-model complexity that Western frameworks have only recently begun to address.

Then there’s the deployment question. China’s Data Security Law (2021) and Personal Information Protection Law (2021) impose strict requirements on cross-border data transfer. The “Xinchuang” (信创) mandate pushes government and state-owned enterprises toward domestic technology stacks. The practical reality of Chinese internet infrastructure means self-hosted models on internal networks are simply more reliable than API calls to international endpoints.

This explains why every major Chinese agent platform — Dify, Coze Studio, Bisheng — treats self-hosted and VPC deployment as first-class citizens, not afterthoughts.

Four Layers of the Stack

The Chinese AI agent ecosystem isn’t a monolith. It spans four distinct layers, each with its own dynamics.

Layer 1: Application Platforms (The Heavyweights)

These aren’t agent frameworks in the LangChain sense. They are full-stack platforms for building, deploying, and managing AI-powered applications.

Dify (langgenius/dify) is the titan. 150,557 stars. 23,721 forks. Created in 2023, pushed to today. Dify is a collaborative workspace for building agentic workflows and RAG pipelines, deployable on cloud, VPC, or self-hosted infrastructure. It occupies the niche that LangSmith + LangServe + LangChain fill in the West, but as a single integrated product rather than a fragmented library ecosystem.

But Dify’s license tells a story the star count obscures. It is technically Apache 2.0, but modified with two restrictions: multi-tenant deployments require a commercial license, and the LOGO cannot be removed. This is the “open-core” pattern perfected by MongoDB and Elastic — open-source for self-hosting, but commercially protected against cloud competitors.

Coze Studio (coze-dev/coze-studio) is ByteDance’s entry. 21,274 stars in just over a year. A visual agent development platform backed by ByteDance’s infrastructure and distribution power. Its rapid adoption reflects the hunger in the Chinese market for production-grade agent tooling.

Bisheng (dataelement/bisheng) is the enterprise-focused alternative at 11,784 stars. Where Dify targets both developers and business users, Bisheng is explicitly enterprise — GenAI workflows, RAG, compliance features, and deployment patterns suited to large Chinese organizations.

Layer 2: Agent Frameworks (The Builders)

MetaGPT (geekan/MetaGPT) is the standout, with 69,565 stars. Self-described as “The Multi-Agent Framework: First AI Software Company,” MetaGPT assigns roles — Product Manager, Architect, Engineer, QA — to different agents and has them collaborate. Its MIT license and ambitious vision have made it the most-starred Chinese agent framework globally.

Qwen-Agent (QwenLM/Qwen-Agent) is Alibaba’s official agent framework at 16,860 stars. It features function calling, MCP support, code interpreter, RAG, and a Chrome extension. It is tightly coupled to Qwen models, representing the integrated model+agent stack pattern common in China. And notably, it lists MCP — Anthropic’s Model Context Protocol — as a feature. MCP adoption is not a Western-only phenomenon.

AgentScope (modelscope/agentscope) is Alibaba DAMO Academy’s contribution, now at 28,353 stars. Its emphasis on “agents you can see, understand and trust” signals a focus on observability and interpretability that goes beyond pure capability.

Layer 3: Autonomous Agents

ChatDev (OpenBMB/ChatDev) at 33,847 stars is the most influential. ChatDev 2.0 frames itself as a virtual software company where agents take on different development roles. Backed by Tsinghua University’s OpenBMB lab, it is a bellwether for Chinese agent research.

UI-TARS (bytedance/UI-TARS) at 11,240 stars represents a distinctly Chinese research priority: GUI automation agents that interact with computer interfaces the way humans do. This reflects China’s mobile-first computing culture — an agent that can navigate a mobile app’s interface is more valuable than one that can write code.

Layer 4: Infrastructure

AIOS (agiresearch/AIOS) at 6,149 stars proposes an “AI Agent Operating System” — a system-level approach to managing multiple agents. Eino (cloudwego/eino) is ByteDance’s Go-based agent framework, bringing production microservice DNA to the agent space. Lagent (InternLM/lagent) at 2,273 stars is Shanghai AI Lab’s lightweight entry.

A Different Design Tradition

Comparing the two ecosystems reveals genuinely different design philosophies — not just localization differences.

Western frameworks are code-first. You write Python, import classes, chain calls. Even “low-code” tools like Flowise are wrappers around code-first frameworks. Chinese platforms are visual-first. Dify, Coze, and Bisheng lead with drag-and-drop workflow builders, visual debugging, and configuration-driven agent creation. The primary user interaction is visual, reflecting a target audience that includes business analysts and product managers, not just developers.

Western frameworks value composability. LangChain’s strength is its hundreds of integrations — swap models, vector stores, and tools independently. Chinese frameworks value integration. Qwen-Agent + Qwen models + ModelScope hub + Alibaba Cloud form a vertically integrated stack. The assumption is that users adopt an entire ecosystem rather than mix and match.

Western frameworks start as research projects. Enterprise features are added later. LangChain’s LangSmith was a post-hoc addition for observability. Chinese frameworks start with enterprise features. Dify ships with user management, audit logs, RBAC, and compliance hooks from the first stable release. This reflects the Chinese market’s emphasis on enterprise and government procurement as the primary revenue path.

The Censorship Layer

All Chinese LLM providers implement content filtering at the API level. Agent frameworks built on top inherit this filtering. But agents present a novel challenge: a multi-step agent might construct queries that pass individual filter checks but produce a problematic output when combined.

This has led to a pattern unique to Chinese agent development: output-level compliance hooks built into the framework itself. Qwen-Agent provides configuration options for content moderation at both input and output boundaries. Dify includes built-in content moderation plugins. Western frameworks do not need this layer and do not have it.

What This Means for the Ecosystem

China’s agent ecosystem is not a copy of Western frameworks adapted for Chinese models. It is a genuinely independent design tradition with its own priorities — visual workflow builders over code-first frameworks, enterprise deployment readiness over research experimentation, and integrated model+agent stacks over model-agnostic tooling.

AllClaws has added Dify, MetaGPT, and Qwen-Agent as tracked platforms, bringing the total to 34. This isn’t about completeness — it is about credibility. A research project that claims to analyze “AI agent architectures” while ignoring 350,000 stars worth of Chinese projects is studying less than half the picture.

The next frontier is interoperability. As MCP gains traction and Chinese frameworks adopt it, the two ecosystems may begin to converge. Whether they converge on a shared protocol, or whether China develops its own standard, is one of the most consequential questions for 2027.

This analysis is based on the full Q3-6 research report: China AI Agent Ecosystem Deep-Dive. Data collected July 2026 via GitHub API.