What exactly did Coze open source this time?

ByteDance's Coze (Button) is no stranger to many of you. Since its launch over a year ago, its features have been continuously refined, and with the popularity of Button Space, it has become a leading product in the consumer-facing workflow application space, helping many users improve efficiency and even generate significant revenue.

Today, ByteDance announced the open-sourcing of Coze's two core projects, Coze Studio and Coze Loop. Together with the previously open-sourced Eino, ByteDance has laid the groundwork for an open-source ecosystem for the development, orchestration, and operations of large-scale AI agents.

Is the open-source version of Coze useful? Why did ByteDance decide to open-source it? What impact will this have on other players in the agent competition? Let me share my personal thoughts.

What exactly did Coze open source this time?
Coze open-sourced two projects today: Coze Studio and Coze Loop.
However, Coze Studio is clearly more popular.

As for the difference between the two, simply put, Coze Studio is the agent building platform that everyone is familiar with, while Coze Loop is the agent operations platform.

Coze Studio: Agent visualization development tool
Coze Studio is the core engine of the platform, providing all the capabilities needed to build agents.

Core capabilities Replicates the online version of the platform, ready to use out of the box, and easy to get started.
Supports intelligent agent/application creation, plugin management, workflow management, database management, knowledge base management, and prompt management.
API & SDK: Supports dialogue and workflow APIs, and can also be integrated into business systems via ChatSDK.

Coze Loop: Agent lifecycle management
If Studio is responsible for “creation,” then Loop is responsible for ‘maintenance’ and “optimization.” It is a solution focused on AgentOps maintenance.

With Loop, developers can turn the mystique of prompts into science:
Full process development: Through a visual playground, developers can intuitively compare the output effects of different models and prompts, and achieve full process management of prompt writing, debugging, optimization, and even version control.

Systematic evaluation: The evaluation module can perform multi-dimensional automated testing of prompts or agent responses.

End-to-end monitoring: Workflow development and troubleshooting is a major challenge, but Loop can record every step from user input to final output, including prompt parsing, model invocation, and tool execution, capturing intermediate processes and abnormal states to make Agent operations transparent.

It can be said that ByteDance's open source release is essentially handing over a “Agent production line + quality control center” that has been verified by a large number of online users to developers.

In terms of technology stack, Coze's backend uses Go, the frontend is React + TypeScript, and the overall architecture is microservices and DDD design, providing a good foundation for secondary development.

Why did ByteDance open-source Coze?
Coze is a commercial platform for end-users, serving tens of thousands of enterprises and millions of developers. Regarding why ByteDance chose to open-source it so generously, the official statement is “committed to lowering the barriers to AI Agent development and application, and encouraging community collaboration and knowledge sharing.” I have two personal speculations.

Accelerate community ecosystem construction and build consensus and standards for agent development

There are many players in the agent field, and the competition for the ecosystem is becoming fierce. Currently, there is no absolute consensus or standard in the industry, and in the future, no single company will be able to define all scenarios.

Before being open-sourced, Coze had already accumulated a huge consumer user base thanks to its extremely low threshold and fun nature.Many talented individuals have developed various practical workflows on Coze and sold their bots or solutions through various channels, earning their first fortune. This proves that Coze's product strength and practicality have been validated by the market.

However, a closed platform may be successful in commercialization in the C-end market, but it cannot meet the requirements of the explosive growth of the B-end ecosystem.

By open-sourcing, Coze significantly lowers the barrier to entry for large-scale model application development, objectively benefiting more enterprises, organizations, and developers.

By attracting more developers to jointly build a rich ecosystem of plugins and applications, they will become familiar with Coze's technology stack; and when enough developers and enterprises develop based on Coze, its development paradigm may become the industry's “de facto standard,” building a moat that is stronger than short-term profits.

Relying on Volcano Engine, Coze has become the infrastructure for Agent Infra

Many people speculate that there will be a “commercial version” of Coze in the future. This is indeed possible. After all, the open source version is responsible for building the ecosystem, while the commercial version is responsible for realizing profits from high-value customers, which is a successful business model that has been repeatedly verified in the open source world.
But I think the scope may still be too small.

Think about it: even if developers master the open-source version to perfection, once they need to deploy it in production environments and serve more users, they will require stable, high-performance computing power, model services, and data storage services. At this point, Volcano Engine, with its seamless integration, is undoubtedly the most cost-effective and least disruptive choice.

In other words, the open source front-end SaaS application is the traffic gateway that attracts developers, while the real commercial value lies in the back-end IaaS cloud services and MaaS model services.

The above is purely my personal speculation, and of course, I don't rule out the possibility that ByteDance will pursue both Coze's cloud services and commercialization at the same time, haha.I think that if we really analyze it from a “big picture” perspective, the prospects for the open source version of Coze are bright, and it will not be a one-off KPI project.

Using the open source version of Coze

For developers, it only takes a few minutes to run Coze Studio, and the hardware requirements are not high (2 cores and 4G are enough), as long as you install docker in advance.
The first step is to download the code repository.
git clone https://github.com/coze-dev/coze-studio.git
Second, configure the model files. The path is backend/conf/model/template/, and configuration files for models like Doubao, OpenAI, Gemini, Claude, DeepSeek, Qwen, and Ollama are pre-configured.

Users can copy the appropriate configuration file based on their actual needs and fill in the available API key.
Step 3: Docker, start!
cd docker
cp .env.example .env
docker compose --profile ‘*’ up -d
After the service starts, visit http://localhost:8888/to open Coze Studio.

Note that you need to enter your email and password the first time you use it, then click “Register”:
The first impression is that it is a simplified version of Koozi:
Upon closer inspection, some modules are indeed missing, such as the UI Builder when creating an application:
(Top: Open-source version; Bottom: Koozi)
For example, when building an agent, the open-source version lacks multi-agent and voice input/output:
(Top: Open-source version; Bottom: Koozi)
Most of the nodes in the workflow are available, but the multi-modal generation part is missing:

The official website also lists the supported features:
Function Module Function Points
Model Service Manage model lists, and connect to online or offline model services such as OpenAI and Volcano Ark
Build Intelligent Agents * Orchestrate, publish, and manage intelligent agents

• Support configuration of workflows, knowledge bases, and other resources Build Applications * Create and publish applications
• Build business logic through workflows Build workflows Create, modify, publish, and delete workflows Development Resources Supports creation and management of the following resources:
• Plugins
• Knowledge bases
• Databases
• Prompts API and SDK * Create sessions, initiate conversations, etc. via OpenAPI
• Integrate agents or applications into your own applications via the Chat SDK

What do you think? I think that if you look at it from the perspective of self-deployment development or complete local deployment + local models, the core components of the open source version of Coze are sufficient for building workflow intelligent agents.

Of course, the more features supported by the official version, the better. Coze's open source version has only just been released, so I hope that more feature modules will be added to the community in the future.

Although it is not perfect, the future looks promising
If I were to evaluate Coze's open-source release, my view is: The current open-source core modules are sufficient to address most scenarios and requirements, but the functionality is not yet perfect (e.g., unable to export applications or workflows from the production environment to the open-source version). However, the future is definitely promising.

Today, I saw some comments in the community, most of which were about the lack of features in the open source version of Coze. However, I think everyone is being too demanding in expecting “absolutely complete and thorough open source.”

Open source ≠ charity. Commercial companies' open source projects usually differentiate between community and commercial versions. Coze Studio has currently open sourced its core engine capabilities, which are already sufficient for developers to build feature-complete workflow agents.

If the open-source community can provide positive feedback and even contribute by actively participating in plugin development, bug fixes, and ecosystem building, and demonstrate the immense potential of the open-source community to the official team, more features may be gradually open-sourced in the future.
I joined the official technical discussion group today, which already has 1,000 members, and found that the voices in the group are very different from those outside. Most of the discussion in the group is about deployment and usage, and I also saw many messages from developers, which were very sincere. Some of the features that everyone strongly requested are already being discussed internally.

Those who argue that “open source is not complete unless it is completely open source” when faced with a project that has just been open sourced for a day are probably not open source contributors.

Open source should not be a unilateral effort by a single company, and the spirit of open source is never about reaping the benefits without contributing. It is about building something together.

This is the beauty of open source: if you think it's not good enough, help make it better.

Where will the workflow agent track go?
Finally, let's talk about competition. When Coze was open-sourced, I discovered something interesting: the user base of Dify and n8n is more active and responsive than that of Coze.

I guess some of the users of Coze don't have a high demand for local deployment, while the user base of Dify and n8n has a great demand for local deployment and even secondary development.

This year, “drag-and-drop workflow agent building” has seen an excellent development opportunity. On one hand, as the model matures, the Agent concept is beginning to take root; on the other hand, DeepSeek all-in-one machines are emerging, and almost all of them will build similar “intelligent agent platforms” to reduce the learning curve for non-technical users through low-code approaches.

After all, the front-end form is “drag-and-drop workflow agent,” so these companies may appear to be competitors.

We won't discuss the functional differences between these platforms today, as that would require tens of thousands of words. However, in terms of open source agreements, Coze is actually the most lenient of the three.

I have summarized the differences in their open-source licenses:
Feature Coze Dify n8n
License Name Apache License 2.0 Modified Apache License 2.0 Sustainable Use License
Commercial Use ✅ Fully permitted
Can be used for any commercial purpose, including resale and building SaaS services.

⚠️ Conditionally allowed
Can be used internally or as a backend for applications. However, multi-tenant services (i.e., a single instance serving multiple tenants/workspaces) are prohibited unless commercial authorization is obtained.
❌ Strictly restricted
Limited to internal business purposes. Direct sale or provision of its core functionality as a paid service is prohibited.
Modification and redistribution ✅ Allowed
Just retain the copyright and modification statements. ⚠️ Restricted
Modification is allowed, but if you use the frontend, you can't remove or modify Dify's logo and copyright information. ⚠️ Restricted

Modification is allowed for internal purposes, but distribution for commercial purposes is prohibited. Removing copyright notices is prohibited.

Core restrictions summary Almost no restrictions, very free. Cannot be used as multi-tenant SaaS, cannot remove front-end brand logos. Cannot be used directly or indirectly for profit, can only be used internally within your company.

Contributor Agreement Follows standard Apache 2.0 contribution rules. Contributors agree that their code may be used for commercial purposes and that the official may modify the agreement. Contributors must sign a CLA authorizing n8n to change the license and use the code for commercial purposes.

In other words, Coze adopts the Apache 2.0 license with no additional terms, allowing free commercial use, secondary development, and the ability to close-source subsequent versions. This effectively imposes almost no commercial restrictions on downstream developers.

These lenient commercial terms are attractive to many who operate outsourcing platforms. Today, I saw a developer joking in the Coze Studio issue tracker: “We used to copy Coze; now we can just use Coze directly.” https://cozestudio.studio/