Build automations with the AI builder

The AI builder is a single surface that is both a chat and a flow editor. You describe what you want, the copilot asks a couple of clarifying questions, and then an AI agent builds the workflow by driving a real browser — step by step, in front of you. When it finishes, the built steps appear on the canvas, ready to review and publish.

Start a build

Open the AI builder (or start a new build from your automations) and describe what you want in a sentence or two. Good prompts mention:

• The trigger — a schedule ("every weekday at 8am"), a webhook, or an app event ("when a new row is added").
• The apps and sites involved — by name.
• The outcome — what should exist or happen when a run finishes.

For example: "Every morning, open my supplier's portal, download yesterday's orders as a CSV, and add each row to my Google Sheet."

Answer a couple of questions

The copilot works against your real workspace, not a generic template. Before it builds, it asks the few things it needs — as pressable option chips or a short dropdown, never a long form. Typical questions: which schedule and timezone, which of your connected accounts to use, or whether a site should run in the cloud or on your own device. You can always type a freeform answer instead of picking a chip.

When it has enough to go on, it shows a Start build card with a summary of what it's about to do.

Watch the agent build

On Start build, an AI agent opens a real browser and works through the task. A progress card shows:

• a live step log ("Opening the portal", "Signing in", "Downloading the CSV") you can minimize,
• screenshots of what the agent sees,
• warnings when a step needs another approach, and
• questions when the agent needs you — for example to complete a login.

You can stop the build at any time; the steps captured so far are kept so you can refine and try again.

Cloud or your device. Builds run in Tyrbo's cloud browser by default. For sites that expect a real, logged-in browser — or that are strict about automated cloud traffic — the copilot recommends running on your device via the Tyrbo Desktop app. If you choose your device and none is connected yet, the chat shows a connect card and waits for one to come online.

Your credentials stay yours. The agent never asks for passwords in chat. To sign in to a site, you connect a login or add credentials to your vault through the secure activation panel — the agent works with the session, never the password.

What a build costs

Unlike the old one-shot builder, an agentic build is metered, not a flat fee — because the agent does real, variable work. A live ticker shows the cost racking up as it works:

• a small starting charge when the agent begins,
• browser time while the agent drives the page, and
• AI tokens for the model's reasoning.

A few things to know:

• Tyrbo checks you have enough credit headroom before letting you press Start build, and every build has a hard cap — the agent pauses and asks before it would exceed it, so a build can never run away with your balance.
• Builds bill for agent time even if you stop them — the agent really did the work up to that point. Stopping early still keeps whatever steps it captured.
• Editing an existing automation with AI stays a flat 50 credits, charged only when you confirm the change — the same as before.

Build costs are separate from run costs: each run of the finished automation debits the normal flat base (5 credits on your device, 10 in the cloud), as with any other automation. See Billing, credits, and overage.

Finish and publish

When the agent is done, the progress card tucks away, a short "assembling your workflow" step runs, and the finished steps appear on the canvas — the trigger, the agent-built browser steps wired to the data they captured, and any code or integration steps. The copilot may run a quick verification (for example, running a generated code step against the scraped sample and showing the output) and ask any last questions, like the schedule's timezone. When it looks right, publish it — after that it runs on schedule like any automation.

Everything is saved as you go, so you can close the tab and come back: reopening the automation restores both the chat history and the flow.

AI steps inside your workflow

The builder can also add run-time AI steps — steps that extract, classify, or summarize data with AI during each run. These are metered separately by token usage. Details and cost controls: Run-time AI steps and credit metering.

Tips for good results

• Name the exact apps ("Google Sheets", not "my spreadsheet tool") so the right integrations are picked.
• One automation per build. Two loosely-related tasks make two clean automations, not one tangled one.
• If a site needs your own logged-in browser, say so — "log in to the vendor portal and download the CSV" tells the agent to plan for a real session.
• Connect the logins and vault credentials a site needs before you start, so the agent doesn't have to stop and wait for you mid-build.