Agent tools available in China

Who this is for

This guide is for people who work in Chinese, or who need a tool with a domestic site, a Chinese interface or domestic model options. It covers five products: Dify, Coze, WorkBuddy, TRAE and Tongyi Lingma. For each, you will see how to start, what the plans cost, which models are offered and what the typical first task looks like. Read the start-up and pricing parts before you commit to a plan.

What the five are

Dify is an open-source LLM app development platform. Its interface combines AI workflow, a RAG pipeline, agent capabilities, model management and observability ^[1]. The README lists hundreds of proprietary and open-source LLMs, including GPT, Mistral, Llama3 and any OpenAI API-compatible models. Its key features are Workflow, Prompt IDE, RAG Pipeline, Agent, LLMOps and Backend-as-a-Service ^[1].

Coze comes in two forms. Coze Studio is the open-source project, an all-in-one AI agent development tool. Its README names prompt, RAG, plugin and workflow as the core technologies, and it is released under the Apache 2.0 license ^[4]. The Coze platform, known in Chinese as 扣子, is a place where people and Agents work together on office, coding, video, file processing and research tasks ^[5].

WorkBuddy is described by its own overview as a full-scenario AI workbench from Tencent. You describe what you need in natural language, and it produces documents, spreadsheets and slides. Its core capabilities run from autonomous planning to local file operations and multi-task parallelism ^[8].

TRAE is a development tool with AI built in. The China site is the Chinese presence of the product, and the same product is also documented in English at the TRAE docs site ^[12][14].

Tongyi Lingma is Alibaba Cloud's coding assistant. Its official site advertises personal free use and free enterprise activation, and it works with Visual Studio Code, Visual Studio and JetBrains IDEs. It also provides its own Lingma IDE ^[17].

Before you start

Check three things before you start: the hardware for self-hosting, the plan and credit system, and the model choices that come with it.

Self-hosting Dify needs at least 2 CPU cores and 4 GiB of RAM. The README also requires Docker Compose v2.35.0 or later, and it licenses the repository under the Dify Open Source License, which is based on Apache 2.0 with additional conditions ^[1].

Coze Studio's quick start needs 2 cores and 4 GB, plus Docker and Docker Compose ^[4].

Several of them bill by credits or message credits. Read the unit before you compare prices. Each of the five has a free tier, but the amount and the reset rule differ.

Dify: self-host or use the cloud

Dify offers Dify Cloud at https://cloud.dify.ai, a self-hosted Dify Community Edition, and Dify Enterprise. The source repository is at https://github.com/langgenius/dify/ ^[1]. The pricing page lists a Sandbox plan at Free with 200 message credits. Professional costs $590 per workspace a year and includes 5,000 message credits a month. Team costs $1590 per workspace a year and includes 10,000 message credits a month. Annual billing saves 17% ^[2].

Message credits let you try models from OpenAI, Anthropic, Gemini, xAI and Tongyi inside Dify. Once they are used up, you can switch to your own API key ^[2]. The Community plan is free, and Enterprise is a custom plan with contact sales only ^[2].

The Chinese homepage describes the Community Edition as open source and self-deployable, and the cloud version as ready to use, for building AI agents quickly. The enterprise edition adds SSO and SAML, RBAC and audit logs ^[3].

To self-host, the quick start runs these commands in order ^[1]:

cd docker
cp .env.example .env
openssl rand -hex 32
docker compose up -d
docker compose ps --all

Then open http://localhost/install in a browser to finish setup ^[1]. Run docker compose ps --all to check that the services are running before you open the page.

Coze: the open-source studio and the platform

Coze Studio's README gives a self-hosted path. It lists model services that integrate OpenAI and Volcengine, and it is released under the Apache 2.0 licence ^[4]. The quick start clones the repository and starts it with Docker Compose ^[4]:

git clone https://github.com/coze-dev/coze-studio.git
cd coze-studio
make web
docker compose -f ./docker/docker-compose.yml up

After it starts, register at http://localhost:8888/sign, then configure models at http://localhost:8888/admin/#model-management ^[4].

The Coze platform itself runs on desktop for macOS and Windows, on mobile and in the browser. The web version needs no installation, but you do need to register and log in ^[5]. Agents can switch between mainstream models such as Doubao-Seed, Kimi, GLM and Minimax, and they are extended with skills, MCP and plugins ^[5].

The personal plans are free, 39.9 yuan, 99 yuan, 199 yuan and 999 yuan a month. The paid personal tiers come with monthly credits of 1,500, 4,500, 10,000 and 55,000 respectively ^[6]. Credits are the only settlement unit on the platform. Coze offers personal, team and enterprise subscriptions. The free personal plan gives activity credits when you log in each day. When the credit balance reaches zero, services stop ^[7].

WorkBuddy: office work and development

The pricing page lists a Trial plan at no charge, with 500 base credits a month. The paid Standard, Advanced and Flagship plans cost 99 yuan, 199 yuan and 999 yuan a month ^[9]. Recurring plans also come with discounts. Under continuous monthly billing (7折), the plans cost 70, 140 and 700 yuan a month. Under continuous annual billing (5.6折), they cost 672, 1,344 and 6,720 yuan a year ^[9]. Top-up packs of 1,000 credits cost 50 yuan and are valid for one month ^[9].

The same page notes that WorkBuddy and CodeBuddy share credits on one account, so you do not subscribe to each separately. The free tier includes real-time code completion at 5,000 uses a month, with the note 限免无限次 (unlimited while the offer is free) ^[9].

The model page lists built-in models such as Hy4 preview, Hy3, GLM-5.3, GLM-5.2, MiniMax-M3, Kimi-K3, Kimi-K2.7-Code, Deepseek-V4-Flash and Deepseek-V4-Pro. It offers three modes: 快速模式, 均衡模式 and 极致模式 ^[10]. Custom model parameters, including your API key, are stored only in the local workbuddy/models.json. Legacy configurations are in ~/.codebuddy/models.json ^[10]. The Tencent Cloud Token Plan supports Hunyuan, MiniMax, Kimi and GLM ^[10].

The scenarios page groups its uses into three areas. Daily office work includes data analysis and visualisation, deep research, and slides. Code development includes websites, WeChat mini programs, Agent applications, Skill development and CI/CD. Design and creative work includes posters, slides, images and video ^[11].

TRAE and Tongyi Lingma

The TRAE China pricing page lists a Free plan at ¥0 with 500 credits a month. The paid tiers are 会员 Lite at ¥49, 会员 Pro at ¥99, 会员 Pro+ at ¥239 and 会员 Ultra at ¥699. Their monthly credits are 2,000, 4,000, 12,000 and 40,000 ^[13]. The page also lists discounts on models including GLM 5.3, Seed-2.1-Turbo, Deepseek-V4.1-Flash and DeepSeek-V4-Pro ^[13].

TRAE's international docs list five plans in US dollars. Its models page includes the built-in models GPT-6.1-Sol, GLM-5.2 and Kimi-K3, among others ^[15]. To add a custom model, open Settings, click Add Model, and enter an API key ^[15].

The billing FAQ says that on-demand usage charges each time the accrued amount reaches $3 ^[16]. The Free plan gets a limited Dollar Usage allowance each month. When it is used up, AI chat is unavailable for the rest of the month ^[16].

Tongyi Lingma has a 编程智能体 mode. It can break a task down, use engineering retrieval, edit files and run terminal tools ^[17]. ## Which one fits which situation

These are short summaries of what each product's own pages say it is for.

• Dify fits teams that want a visual workflow and RAG tool, either as a cloud service or self-hosted with Docker ^[1][2].
• Coze Studio fits people who want an open-source agent builder they can run themselves under Apache 2.0 ^[4]. The Coze platform fits people who want to work with agents across desktop, mobile and web ^[5].
• WorkBuddy fits office work that starts from a plain-language request, such as documents, a spreadsheet or a slide deck ^[8][11].
• TRAE fits developers who want an IDE with a separate Agent window. The China plans and the international plans differ in price and currency ^[13][15].
• Tongyi Lingma fits developers who already use Visual Studio Code, Visual Studio or a JetBrains IDE, and who want an assistant inside it ^[17].

Limits and common questions

• Free tiers run out. Dify's Sandbox plan is capped at 200 message credits. Coze stops services when the balance reaches zero. TRAE's Free plan stops AI chat when its monthly allowance runs out ^[2][7][16].
• Prices differ by region. TRAE's China plans are in yuan, and its international plans are in US dollars ^[13][15].
• Self-hosting is a server job. Dify and Coze Studio both need Docker, and Dify needs at least 2 cores and 4 GiB of RAM ^[1][4].
• Model choice depends on the plan. On TRAE's Free plan, model selection is limited to Auto mode ^[15][16].
• Model and price lists change. Check each vendor's pricing or models page before you commit to a plan ^[9][13].

Next steps

For Dify, start with the README and its quick start ^[1]. For Coze Studio, read the README's quick start ^[4]. For the Coze platform, read the docs ^[5][6]. For WorkBuddy, start with the overview and the pricing page ^[8][9]. For TRAE, read the About page and the plans FAQ ^[14][16]. For Tongyi Lingma, start with the official site ^[17].