Migrate from Coze to LoopEngine

ByteDance is folding Coze into Doubao, and Coze's business plans sold through Volcano Engine end on 31 December 2026, with no export path announced yet. This guide moves a Coze bot to a LoopEngine agent: what to save from Coze first, how each Coze feature maps to LoopEngine, and the steps to switch over without a gap.

LoopEngine agents are plain files in your workspace, and the runtime's source is available to run yourself. If you ever leave LoopEngine, you take your agents with you, or run them on your own servers.

1. Save what you need from Coze

Do this first, while your Coze account still works. Nothing here needs an export feature: it's copying and downloading. For each bot, keep:

• Persona and prompt: copy the text into a document.
• Model and settings: which model it uses, and anything notable like temperature.
• Plugins: the list, and for custom plugins their API endpoints, parameters and the credentials they use.
• Workflows: a screenshot of each canvas, plus a few lines describing what it does in plain words: what starts it, each step, where it branches, what it returns. This description is what you'll give LoopEngine.
• Knowledge bases: the original files (or where they came from). Uploading the originals again is easier than extracting them from Coze.
• Triggers: each scheduled task and its timing.
• Channels and callers: where the bot is published (Feishu, Slack, a website) and any systems calling it through the Coze API.
• Test conversations: five or ten real questions with answers you were happy with, to check the new agent against.

2. How Coze maps to LoopEngine

The biggest difference is how workflows run. A Coze workflow follows the graph you drew, node by node. A LoopEngine agent follows its instructions with the tools you give it, and decides the next step from what it finds. Instead of drawing every branch, you decide what it may do, tool by tool:

• Allow: it runs on its own.
• Ask me: it waits for a person's approval.
• Never: it's blocked.

3. Set up a workspace

1. Sign up and create a workspace from your dashboard.
2. Open Models in the workspace and save a key for the provider you want. If you used Doubao on Coze, DeepSeek, Kimi and GLM are available alongside OpenAI, Anthropic and Gemini.
3. Open Apps and connect the services your bots used, with their API keys.

4. Recreate each bot as an agent

1. Check the templates first: if one fits the bot's job (answering customer emails, refund approvals, incident triage, expense checks and more), start from it. Otherwise choose + New next to Agents.
2. On the agent's Overview tab, paste the bot's persona and prompt as its identity, and pick its model.
3. On its Apps tab, add the apps it needs and tick only the tools this agent uses, the same way you chose plugins for the bot.
4. On its Permissions tab, set anything that changes data, spends money or contacts a customer to Ask me. Coze bots have no per-tool approval setting, so this is the step most worth your time.
5. On its Messaging tab, choose where approval requests arrive: Slack, Lark, email, Telegram or a webhook.

5. Rebuild workflows

Open the agent from step 4, choose Build a workflow on its Workflows tab, and paste the plain description you wrote in step 1, for example: "Every morning at 9, read new orders over $500 from Shopify, check each customer's history, and post a summary in Slack. Ask me before cancelling any order." LoopEngine plans the steps, shows which ones ask you first, lists any app or messaging the agent still needs, and adds the workflow to that agent. See Workflows.

How the usual Coze nodes carry over:

• Model (LLM) nodes: the agent itself. Its instructions say what to write or decide.
• Plugin nodes: app or ability tools on the agent.
• Condition nodes: plain rules in the instructions ("if the order is over $500…"), and an approval wherever a person should decide.
• Code nodes: usually unnecessary, since the model can reformat and calculate. For exact logic, a developer writes a small tool and adds it as a private ability. See the abilities docs.
• Knowledge nodes: a search tool on the agent: Docs search for your files, or the app where the documents live.
• Nested workflows: a separate agent for the sub-job; developers can make one agent a tool of another.

6. Schedules and callers

• Scheduled tasks become schedules on the agent's Schedules tab. They start paused; turn them on once you've tested the agent.
• Systems calling the Coze API call the agent instead. Create an API key on your workspace's page in the dashboard, then:

POST https://<your-workspace>.loopengine.app/agents/<agent-name>/messages
Authorization: Bearer lek_...
Content-Type: application/json

{ "message": "Where is order 1001?" }

The reply is in the response's text. If the agent stopped to ask for an approval, the response is 202 with pending: true.

7. Test, then switch over

1. Ask the new agent your saved test questions in its Chat tab and compare the answers with Coze's. Adjust its instructions until you're happy.
2. Run both side by side for a few days if you can, with the agent's risky tools on Ask me.
3. Point your channels and API callers at LoopEngine, and turn on the agent's schedules.
4. Turn off the Coze bot's triggers so nothing runs twice, and keep your saved copies until the Coze plan ends.

What works differently

• No canvas. You describe workflows instead of drawing them, and the agent adapts to what it finds rather than following fixed branches.
• Approvals are built in. Risky actions wait for a person in Slack, Lark, email or Telegram, and the wait survives restarts.
• Each agent gets only the tools you give it, with a permission on each one, rather than a whole plugin.
• No public bot store or one-click publishing to chat platforms. LoopEngine agents work for your team and your systems; for a customer-facing chat, call the agent from your own site through the API.
• Your agents stay yours. Run them in LoopEngine Cloud or on your own servers; see For developers.

Questions about moving over? See the Help center, or open an issue on GitHub.