AI assistant overview
What it can do, where to open it, how to configure the model, plus the slash commands and page shortcuts that cost no quota.
The AI assistant is not just a YAML generator. It reads the context of the page you are on — the open workflow, the selected step, the run you are looking at — so it can build flows, investigate failures, launch jobs and change settings.
Where To Find It
- the floating button in the bottom-right corner, present on every main page;
- Shortcuts
Alt + G; - the "hand it to the AI" entry points on the editor's diagnostics panel and step properties panel — going in this way carries the current context along automatically.
Six kinds of things it can do
| Category | Example |
|---|---|
| Build workflows | Build a flow from scratch, change one step of an existing one, draft from a template, validate and save. |
| Diagnose failed runs | "Why did this run fail?", "fix it and run it again for me". |
| Recommend templates and modules | Describe your business goal and let it search existing templates and official modules first, rather than inventing something. |
| Scheduling and in-app actions | Create a schedule, enable or pause one, run or validate the current workflow, cancel a run, refresh skills, switch the interface language. |
| Navigation and settings | "Take me to schedule management", "where do I change this setting?". Pure navigation requests skip the model entirely and land instantly at no cost. |
| Answer product questions | How to write a given field, whether a capability is supported, how to call a given API. |
Set it up once
Before the first use, go to Account Settings → AI Assistant turn it on and pick a model provider (OpenRouter, OpenAI, Anthropic, DeepSeek and so on), or configure a subscription model in the Claude Code / Codex section. The secret itself lives in Credential Management → My Credentials, and at run time it resolves against the current account — administrators do not need to pre-configure anything for ordinary users.
Three runtime identities
The assistant itself can be the built-in Koog engine, or the Claude Code or Codex command line — switch it in the first field, assistant identity type, under Account settings → AI assistant. External identities need the matching subscription credential, and whether they can edit workflows in the conversation depends on the deployment configuration; if you mainly want help building flows, staying on Koog is simplest.
Switch the assistant's identity and model
The AI assistant discussed here uses your own account's assistant configuration — it is you talking to the platform. The globalAgent field at the top of a workflow's YAML is that workflow's own agent configuration at run time. The two do not affect each other.
Slash commands and quick actions
Commands typed with a leading slash call platform endpoints directly without going through the model, so they return instantly and consume no assistant quota:
/new 新建工作流
/execute 执行当前工作流（别名 /run）
/validate 校验当前工作流
/export 导出当前工作流 YAML
/undo 撤销 AI 对工作流的上一次修改
/help 列出可用命令Above the input box you also get shortcut buttons for the page you are on: run / validate / export / undo in the editor, diagnose / retry on an execution detail page, new / create from template on the workflow list, and view recent failures on the execution list.
Read next
- Switch the assistant's identity and model — How to choose the identity, where to change the model, and how to clear the errors.
- Build workflows with it — Diff preview, undo stack, canvas sync, and template and module search.
- Diagnose failed runs — Hand it a failed run to locate, explain and fix.
- Sessions, memory, and background runs — Managing several sessions, long-term memory, and tasks that keep running after you close the window.