Skip to Main Content
AI Assistant

Say what you need — it builds the workflow, runs it, and fixes it when it breaks

The AI assistant is more than 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, debug errors, launch runs, and change settings.

Six task typesAvailable everywhere · Alt + GThree Runtimes
Common uses

Every role opens with a different first sentence

Operations"Move this daily report to 9am and message me if it fails"
Campaigns"Spend jumped 40% yesterday — find out which campaigns"
Engineering"Add a code step here that calls our internal API"
Team leads"Which workflows cost the most this month? Make me a table"
How it works

You describe the work, it edits the canvas, and every change waits for your confirmation

Every change arrives as a diff: which steps were added, which fields changed. Nothing is written to the workflow until you hit save — undo it and the canvas rolls back with it.

  • It searches existing templates and official modules first, and only generates from scratch when nothing fits
  • Whatever it generates is validated immediately, so definition errors surface right away instead of at run time
  • The canvas follows along: whichever step it is talking about is the step highlighted
Build workflows with it
AI AssistantAlt+G
You

Every morning, pull yesterday's ad data; if spend looks off, ask me before adjusting the budget.

Assistant

Found a close template, "Daily ad check", and added an approval gate on top of it. Here is the diff:

- step: fetch_spend - step: detect_anomaly+ - step: adjust_budget+ manual_approval: {enabled: true}~ schedule: 每天 08:00
SaveCancelRun it once
You

Why did last night's run fail?

Assistant

Step 3 failed authentication because its credential expired. I have located the exact field; once fixed, this run can be retried directly.

View diagnosisFix and retry
What it can do

Six task types, from building flows to chasing failures

It is not confined to the editor: the floating button sits on every main page, and the "Ask AI" entry points in the diagnostics and step property panels carry the current context with them.

01Build workflowsBuild a flow from scratch, change one step of an existing one, or draft from a template — validated before saving.
02Diagnose failed runsHand it a failed run: it locates the cause, explains it, fixes it, and runs it again.
03Recommend templates and modulesDescribe the goal and it searches existing templates and official modules before inventing something new.
04Scheduling and in-app actionsCreate scheduled jobs, start or stop them, run or validate the current workflow, cancel a run, switch the interface language.
05Navigation and settings"Take me to schedules", "where do I change this setting" — pure navigation skips the model and costs nothing.
06Answer product questionsHow to write a field, whether a capability exists, how to call an API — answered from a documentation search.
Zero-cost path

Commands and navigation skip the model entirely

Commands that start with "/" call the platform API directly: instant, and they never touch your assistant quota. Pure navigation requests skip the model too — "take me to schedules" is a jump, not an inference.

  • Quick actions above the input adapt to the page — in the editor: run, validate, export, undo
  • A run detail page offers diagnose and retry; the run list surfaces the latest failures right away
  • On the workflow list it is create and create-from-template, a few menu levels saved
Slash commandsNo quota used
  • /newNew Workflow
  • /executeRun the current workflow (alias /run)
  • /validateValidate the current workflow
  • /exportExport the current workflow YAML
  • /undoUndo the assistant's last change to the workflow
  • /helpList the available commands
Runtimes

Three engines to choose from, and subscriptions work by signing in

Turn it on and pick a model provider under Account settings → AI assistant. Keys live in Credentials → My credentials and are resolved per account at run time, so an admin never has to pre-configure anyone.

KoogAPI Key
The built-in runtime. Connect OpenRouter, OpenAI, Anthropic, DeepSeek and other providers — one key is enough.
Claude CodeSubscription login
With a Claude subscription you can sign in to drive the assistant — no separate API key needed.
CodexSubscription login
Same for a ChatGPT subscription: sign in and go, and let it finish long-running tasks on its own.
Sessions and memory

Close the window, the task keeps running

Parallel sessionsOne session per workflow, with no cross-talk; switch back any time and pick up the earlier context.
Long-term memoryYour usual style, preferred models, and naming conventions are remembered, so you do not explain them again.
Background runsClose the assistant window or even the page — the running task carries on and the result is waiting when you return.
FAQ

The questions people ask before they start

Boundaries

Three things to be clear about upfront

  • The assistant edits but never saves: every change waits as a diff until you confirm and hit save. No silent writes.
  • It runs on your own account's assistant settings, which are not the same thing as the workflow YAML's top-level globalAgent. The two do not affect each other.
  • Model calls go through the provider and credentials you configured and are billed at that provider's rates; slash commands and pure navigation consume no quota.
Hand the assistant your first request
The Free plan includes unlimited assistant use with your own credentials — and the editor is always there when you would rather do it yourself.