Coze
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,360 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,360 words
Add missing citations, update stale details, or suggest a clearer explanation.
Coze is a no-code and low-code platform from ByteDance for building AI agents and chatbots. Users assemble a bot through a visual builder by writing a prompt that defines its persona and instructions, then extend it with plugins (external tools and APIs), multi-step workflows, knowledge bases for retrieval over uploaded documents, variables, databases, and scheduled tasks. Finished bots can be published to channels such as a hosted web page, an API, and messaging apps including Discord, Telegram, Slack, and ByteDance's own Feishu/Lark. The platform launched in 2023 and is widely described as ByteDance's answer to OpenAI's GPTs.[1][2]
Coze runs in two largely separate forms: the international service at coze.com, and a Chinese version branded Kouzi (扣子) operated through ByteDance's cloud arm, Volcano Engine. In July 2025 ByteDance open-sourced the core of the platform as Coze Studio, together with a companion observability and evaluation tool called Coze Loop, both under the Apache 2.0 license.[3][4]
The product targets people who want to build a conversational AI application without writing much code. A creator starts a bot, gives it a system prompt, picks an underlying large language model, and attaches capabilities. Beyond simple single-prompt bots, Coze supports a multi-agent mode in which several specialized agents are connected and the conversation is routed between them, and a workflow mode for tasks that need fixed, repeatable logic rather than free-form generation.[1][5]
Coze is frequently compared to OpenAI's GPTs (custom versions of ChatGPT) because both let non-programmers package a model with instructions and tools. It is also compared to Dify, an open-source LLM application platform, particularly after Coze Studio was itself open-sourced.[3][6]
The builder is organized around a set of reusable components that a bot or workflow can call.
| Component | What it does |
|---|---|
| Prompt and persona | Defines the bot's role, tone, and instructions in natural language. |
| Plugins | Tools that call external APIs (for example web search, image generation, or a GitHub lookup); creators can also register custom plugins from their own APIs. |
| Workflows | Visual, node-based pipelines that chain models, plugins, code blocks, and logic for complex or high-reliability tasks. |
| Knowledge bases | Document stores (PDF, TXT, DOCX, web pages, and others) that the bot retrieves from to ground its answers, an application of retrieval-augmented generation. |
| Databases | Structured tables the bot can read from and write to during a conversation. |
| Variables | Persistent values, used for things like remembering user preferences across sessions. |
| Tasks | Scheduled or triggered actions, such as sending a daily news summary at a set time. |
Plugins are central to the experience: the platform ships a library of ready-made plugins and lets creators turn an existing API into a plugin by describing its parameters. Workflows let creators combine plugins, model calls, and code in a drag-and-drop canvas, which is useful for processes that must run the same way every time rather than relying on the model to improvise.[5][6]
Once built, a bot can be published to multiple destinations from a single configuration, including a Coze-hosted web page, an embeddable chat widget, an API and SDK for custom integration, and third-party channels such as Discord, Telegram, Slack, and LINE, alongside ByteDance's Feishu/Lark.[2][6]
The two editions of Coze run on different infrastructure, serve different regions, and draw on different underlying models. The international site (coze.com) was introduced for overseas users in 2023 and offers a choice of third-party foundation models. The Chinese version, Kouzi (扣子), launched in early February 2024 and is built around ByteDance's in-house Doubao models; OpenAI's services are not officially available in mainland China, so a domestic model stack is required.[1][7]
| Aspect | International (coze.com) | China (Kouzi / 扣子) |
|---|---|---|
| Operator | ByteDance | ByteDance via Volcano Engine |
| Launch | 2023 | February 2024 |
| Underlying models | Choice of third-party models (for example OpenAI GPT-4 family and Google Gemini) | ByteDance Doubao models |
| Primary channels | Web, API/SDK, Discord, Telegram, Slack, LINE | Web, API, Feishu, and other domestic surfaces |
The China launch was covered as part of a broader ByteDance generative-AI push, with the company positioning Kouzi against ChatGPT and GPTs in a market where ChatGPT is not officially offered. ByteDance's general-purpose Doubao models support a large family of its consumer and developer products, of which Kouzi is one.[1][7]
A defining difference between the two editions is model choice. On coze.com, creators can select among external foundation models and pick one per bot to trade off cost, speed, and capability; coverage reported by users and third-party guides has included OpenAI's GPT-4 and GPT-4 Turbo, GPT-4o, and Google's Gemini, among others, with the exact lineup changing over time. The Kouzi version instead uses ByteDance's own Doubao model family, the same general-purpose and specialized models that power Doubao, ByteDance's consumer chatbot.[2][8]
On 25 July 2025, ByteDance open-sourced two core projects from the Coze stack: Coze Studio and Coze Loop. Both were released under the permissive Apache 2.0 license, which allows free use, modification, and commercial deployment.[3][4]
Coze Studio is the open-source edition of the agent-development core that powers the hosted platform. ByteDance described it as derived from the same Coze Development Platform that had served tens of thousands of enterprises and millions of developers. It provides a visual, no-code or low-code environment for creating, debugging, and deploying agents, applications, and workflows, with the same building blocks as the hosted product: a drag-and-drop workflow engine, plugins, knowledge bases with RAG, databases, prompt resources, and publishing through an API and Chat SDK. Model access is configured by the operator: the project documents connecting to model services such as OpenAI and Volcano Engine's Ark, rather than shipping a fixed model. The backend is written in Go and the frontend in React with TypeScript, organized as microservices following domain-driven design; it can be deployed with Docker Compose.[3][9][10]
Coze Loop is a companion platform for the operational side of agent development, covering prompt development and debugging, automated evaluation, and monitoring across an agent's lifecycle, with SDKs in multiple languages. The two projects rely on Eino, ByteDance's open-source Go framework for orchestrating LLM applications, which had been released earlier in 2025.[4][10]
The release drew rapid attention on GitHub, accumulating roughly 9,000 stars across the two repositories within about 48 hours, with Coze Studio accounting for most of them. As of early 2026 the Coze Studio repository continued to be developed, with tagged releases published over the following months.[10][3]
Coze has been received as an accessible on-ramp to building AI agents, especially for users who want GPT-4 class models and a plugin and workflow system without coding, and for a time it was notable for letting people use capable models at no upfront cost. Early Chinese-market testing was more mixed, with some reviewers finding the initial Kouzi release short of expectations relative to international tools. The 2025 open-sourcing of Coze Studio under Apache 2.0 was widely read as a competitive move in the crowded agent-platform space, where it is measured against both proprietary offerings such as OpenAI's GPTs and open-source projects such as Dify.[3][6][11]