Tabductor
Tabductor
Sign inGet started
Guide

Writing a workflow

A workflow is a set of instructions in plain language: what to do, on which site, and what to bring back. You write it once; Tabductor runs it whenever you, a schedule, an email or the API asks.

Write the instructions

On Workflows, describe the whole job and choose Create workflow. The workflow is named after the first line; rename it in the editor. Good instructions say:

  • Where: the site or page to start from.
  • What to do: the steps, in the order a person would take them.
  • What to bring back: the details you want, and where to keep them (the result, the store, files).
  • Constraints: what not to touch, and what counts as done. “No matching results” is a valid outcome, not a failure.
Every Monday, open the vendor portal, go to Invoices, filter to last week,
download each invoice PDF and save it, and add each invoice to the store
with its number, date, vendor and total.

Your instructions are the only thing the workflow takes orders from. Text on the websites it visits, and values passed in as inputs, are treated as data.

Inputs with $variables

Write $name wherever a value should be filled in at run time, for example Search for flights to $city on $date.

  • A name starts with a letter or underscore, followed by letters, digits, underscores or hyphens. Using the same name twice means one value.
  • Every input is required and can't be blank. There are no defaults: put fixed values in the instructions instead.
  • A dollar sign followed by a digit, like $20, is just a price. Write $$ for a literal dollar sign before a letter.

Where the values come from:

  • Run workflow in the editor shows a field per input. These values aren't saved with the workflow.
  • Each schedule has its own values.
  • The API takes them as inputs, named without the $.
  • An email can fill only $email_from, $email_subject, $email_body and $email_files.

Choose what it returns

Without anything else, a run returns a written summary. To get structured data back, open Choose what it returns and pick One answer with these details or A list of items, each with these details, then add each detail with a name, a type, optional notes for the AI, and whether it must always be included. Types: Text, Number, Whole number, Yes / No, Date, Web link, One of a few choices, List of text, List of records. You can also edit the result as JSON Schema.

Passwords and other secrets

The simplest way to stay signed in is a profile. When a run must type a password, card number or API key itself, add it under Settings → Secrets with a name, the websites it may be used on, and the value. Mention the secret by its name in your instructions, such as “sign in with the Bank login password”. Don't write it as a $variable: that makes it an input. The agent fills a secret into the page without ever seeing its value, and only on the websites you allowed.

Saving and versions

Saving is publishing: each save that changes the instructions or the result creates a new version, and new runs use it. Run workflow saves first, then starts a run. Changing only schedules keeps the current version.

From agent to script

The first run of a version is done by an AI agent working in a real browser. When that run succeeds cleanly, Tabductor distills it into a script, so the next runs repeat the same steps in seconds, without paying for a model. The editor shows what the next run will use:

  • AI agent: the next run uses the agent. A clean successful run will be distilled.
  • Analyzing run / Compiling: the script is being built from the latest run.
  • Distilled · v<n>: the next run uses the script, and falls back to the agent only if the page changed. After a fallback the script is rebuilt from that run.
  • Not compiled, with the reason: for example the last run didn't succeed, or needed a person. The next clean run tries again.

Changing the instructions starts a new version that runs with the agent until it is distilled again.

Choosing the model

Agent runs use your account's default model, set under Settings → Models. To use a different model for one workflow, pick it in that workflow's Settings under Workflow model, or choose Account default to go back. Each model lists its rate; script runs use no model.

Running it

Choose Run workflow in the editor. A run needs a browser profile and enough balance; if either is missing, the editor says what to fix. Then follow the run on the Runs tab.

Writing a workflow · Tabductor docs