Coze
also known as 扣子, Coze Studio
Type: low-code · Vendor: ByteDance (字节跳动) · Language: N/A · License: proprietary (hosted); Coze Studio open-source edition is Apache-2.0 · Status: active · Status in practice: mature · First released: 2024-02-01
Provide a hosted, no-code visual platform on which a non-engineer can compose a bot from large-language-model nodes, plugins (tools), a knowledge base (RAG), workflows, memory and triggers, and publish it to chat channels such as Discord, Telegram, Lark/Feishu and WeChat.
Description. Coze is a no-code/low-code agent-building platform from ByteDance, launched internationally on 2024-02-01 and operated in China as 扣子 on coze.cn, where it is powered primarily by ByteDance's Doubao (豆包) models. A creator assembles a bot from four reusable building blocks — a workflow graph, plugins (callable tools), a knowledge base (RAG over uploaded documents) and memory/variables — in a drag-and-drop editor, then publishes the bot to one or more chat channels (Discord, Telegram, Slack, Lark/Feishu, WeChat). Multi-Agent Mode lets several specialised bots be wired together with jump conditions, so a 'main menu' bot can route a turn to a specialised sub-bot. The open-source Apache-2.0 edition (Coze Studio) is published by ByteDance's coze-dev GitHub organisation and derives directly from the same Coze platform.
Agent loop shape. Visual workflow graph composed of typed nodes — model, plugin (tool), knowledge-base retrieval, conditional/code — driven from a Start node and terminated by an End node. A bot binds a workflow (or a single Agent prompt) to plugins, knowledge bases, databases, and memory/variables. Bots can also be wired in Multi-Agent Mode with jump conditions, and invoked by user messages on published channels, scheduled triggers, or webhook event triggers.
Primary use cases
- no-code chatbot building published to Discord/Telegram/Slack/Feishu/WeChat
- RAG over uploaded documents via the Knowledge Base resource
- scheduled and webhook-triggered bot tasks
- multi-bot routing via Multi-Agent Mode
- self-hosted agent development via the open-source Coze Studio edition
Key concepts
- Bot (智能体) (docs) — The publishable unit — prompt + plugins + knowledge + workflow + memory.
- Workflow (工作流) → visual-workflow-graph (docs) — Drag-and-drop graph of nodes running a multi-step process.
- Plugin (插件) → tool-use (docs) — Registered tools (HTTP/API endpoints or built-ins) the bot can call.
- Knowledge Base (知识库) → agentic-rag (docs) — Uploaded documents indexed for retrieval; queried from a knowledge-base node inside a workflow.
- Memory / Variables (记忆 / 变量) (docs) — Per-user history and key/value variables persisted by the bot.
- Trigger (触发器) → event-driven-agent (docs) — Scheduled cron and event/webhook triggers that invoke a bot or workflow.
Patterns this low-code implements —
- ★★Agentic RAG
Knowledge Base resource + dedicated knowledge-base retrieval node in workflows. Supports hybrid search (full-text + semantic) and rerank.
- ★★Event-Driven Agent
Event-trigger generates a webhook URL; HTTPS POST to the URL invokes the bot/workflow. Bots are also published to chat channels where incoming messages are events.
- ★★Scheduled Agent
Timed triggers support 'every day', 'working day', 'every week X', specific YYYY-MM-DD dates, and full cron expressions.
- ★★Supervisor
Multi-Agent Mode wires specialised bots with jump conditions; the 'main menu' bot acts as a central router that delegates to and returns from sub-bots.
- ★★Tool Use
Plugin system: a registered API endpoint becomes a callable tool the agent can invoke. Built-in and custom plugins both supported.
- ★★Visual Workflow Graph
Workflow editor is described in the Coze Studio README as a drag-and-drop visual canvas for assembling typed nodes into an executable business process.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.