Research
Find a box photo for Wingspan and put it on the product.
- search_images · “wingspan board game box”
- import_image → product
Imported the second result — front of the box, 1600 px, no watermark.
Chat Workflows Approvals Costs you can see
Agents that do the work.
And show their work.
Two ways to put Clear's AI to work. Ask the assistant in a chat, or draw a workflow that runs by itself — on a schedule, when an order is paid, when another system calls. Both use the same tools, act with your permissions, and ask before they change anything that matters.
A chat that has hands.
The assistant sits beside every screen of the panel. It does not describe what you could do — it does it, in your account, and shows you each step.
Pictures
Make a hero picture: board games on a wooden table, warm evening light.
- generate_image · $0.04
Here it is, saved to your media library. Want it on the home page hero?
generated · 1536 × 672
In place
Home › hero_banner › headline
Make this shorter.
- update_component · preview
“Six hundred games. Good coffee.” — nothing is changed until you confirm.
ConfirmCancel
Tools across the whole product
Pages and components, catalog, orders, customers, coupons, bookings, services, media, SEO, marketing — and workflows themselves. If there is a screen for it, there is a tool for it.
It knows where you are
The page, component or workflow you have open travels with your message. “Make this shorter” means this.
Every step, and its price
Tool calls are listed as they happen, and each reply carries what the model and its tools cost — per step and per conversation.
It never locks
Keep typing while it works. Messages queue and go out in order; attach a picture or a spreadsheet to any of them.
Draw it once. It runs itself.
One run of one workflow, start to finish. The canvas shows where it ended; the steps below are how it got there.
01Published, and waiting
An event trigger is switched on by a person. Until an order is paid, nothing runs and nothing is billed.
02An order is paid
The run starts on the published version and pins it — editing the workflow now cannot change this run.
03The condition picks a branch
$212 is over $150. The other branch is marked skipped, not left hanging.
04An agent step does the thinking
A real agent with the tools you bound to it and a step budget. Its tokens and cost are on the node.
05The write waits for you
A step that changes something previews first and parks. The run holds here — for days, if it has to.
06Approved, written once, done
The note is on the order. The run closes with what each step took and what it cost.
Four ways to start a run.
A workflow can have more than one. The ones that fire by themselves only ever run the version you published — never the draft somebody is halfway through editing.
By hand
Declare the questions a run needs — a percentage, a category, a website — and Run shows them as a form. Dry run first, if you like.
DirectionDown
Percent10
Dry runOn a schedule
Pick the days and the hour in your own time zone. A slot that falls while the last run is still going is skipped, and recorded.
MTWTFSS
At08:00 · America/Los_Angeles
When something happens
An order is paid, stock runs low, a form comes in. Add filters over the event, and use its fields in any step.
Eventlisting.low_stock
Only whenquantity ≤ 2
From another system
Each webhook trigger gets its own secret address, shown once. Optional signature check; the caller gets the run's id back.
POST /hooks/workflows/••••••
→ 202 { run_id }
→ 202 { run_id }
- order.paid
- order.shipped
- order.delivered
- order.cancelled
- listing.low_stock
- listing.out_of_stock
- booking.created
- booking.cancelled
- form.submitted
- subscriber.created
- customer.created
- review.submitted
- donation.paid
- workflow.run_failed
Nine parts. That is the whole language.
Steps pass their results along by name — {{ nodes.big.total }}. A reference to something that does not exist is caught when you save, not at three in the morning.
TriggerStart
Where a run begins, and the questions it asks.
ToolActions
One thing the platform can do — the very tools the assistant uses, with the same forms.
AgentAI
A model with the tools you hand it, a step budget, and an answer in the shape you ask for.
ConditionLogic
Rules joined by and / or — or let the model sort things into named branches.
For eachLogic
A list in, one step per item, several at a time. One bad item does not end the rest.
TransformData
Filter, map, sort, limit, do the arithmetic — no model involved, so no cost.
ApprovalLogic
Stop and ask a person. Approve goes one way, reject the other.
WorkflowActions
Run another workflow as a step. Build small ones and compose them.
OutputData
What the run hands back — to you, to the assistant, to the workflow that called it.
Two to start from:Reprice listings — a percentage or an amount, up or down, to a price ending, by category.Re-theme a website — a new look: colours, banners and pictures, made and applied.
It can never do more than the person behind it.
Automation that touches prices, stock and customers has to be boring about safety. These are rules in the engine — not sentences in a prompt that a clever request can talk its way around.
It runs as a person
The one who pressed Run, or who switched the schedule on. Permissions are read at every step: lose the right to edit listings, and that step is refused.
Writes preview first
A step that changes something shows what it matched and the before and after, then parks. Approve writes it; reject sends the run down the other path.
Money is never automatic
You may let a trusted step approve itself. Refunds, shipping labels, buying a domain, launching an ad or raising its budget cannot be set that way.
Some things only people do
Approving, publishing, switching a schedule on: none of these is a tool. The assistant can build and test a workflow; it cannot let it loose.
A run keeps its version
Every run pins the graph it started with. Editing the workflow never changes — or breaks — a run that is already going.
A budget per run
Set the most a run may spend. When it gets there it stops, keeps what it produced, and can be resumed once you have topped up.
Or skip the canvas and ask for one.
The assistant reads the same guide the editor is built from, draws the whole graph in one go, checks it, and dry-runs it. You open it on the canvas already laid out.
Create a workflow that takes 10% off every puzzle on the first of each month, and ends prices in .99.
- get_workflow_guide
- list_workflow_tools · catalog
- create_workflow · 5 nodes
- validate_workflow · no problems
- run_workflow · dry_run
Built from the Reprice template and dry-run: 23 listings in Puzzles would change, nothing was written. Two things are yours to do — publish it, and switch the monthly schedule on. I can't do either.
Run receipt
completed · 3.2 s- Which listings, and by how muchtrigger—
- Change the pricestool · preview—
- Write the summaryagent · 1,480 tokens$0.015
- What changedoutput—
From your balance$0.02
Only thinking costs
Steps that use a model, or make a picture, are metered. Conditions, transforms, loops and ordinary tools cost nothing to run.
Priced where you look
Cost per step and per run in the editor, the last and the average beside the Run button, an estimate from a dry run, and a Workflows filter on your balance.
Limits by plan
How many workflows you keep and how many live runs a month come with the plan; dry runs do not count. See the plans.
Ask for a website.
Bring a logo, a product sheet, a rough idea — or nothing at all. The first fourteen days are free.