What the agent does, what you do, and when.

Ask in a sentence

Describe what you want from the internet the way you would tell a colleague. Add URLs or search criteria if you have them. You do not need to know how the source works.

• The agent turns the sentence into explicit sources, fields, filters, volume, and output.
• You revise before approving a run.
• The agreed scope is stored with the job record.

A straight answer: possible or not, and for how much

Before you pay, an agent reads the sources and reports back: which sites it can reach, which fields exist, what will be empty, and what it costs.

• Sources that block automated access or hide data behind a login are named up front.
• When something is out of reach, you are offered what can be collected instead.
• Your request is never quietly swapped for an easier one.

This step is the product. Technical people already know almost anything on the web is reachable. The agent makes that knowledge available to everyone, one request at a time.

User-defined columns

Select, add, remove, reorder, rename, and describe columns. Give each one a type: text, number, URL, date, currency, or boolean.

• Each column is labeled collected, derived, or extra work.
• The estimate updates when a column adds paid requests.
• Missing values stay missing. Inferred fields say how they were produced.

Renaming product_title to Product name triggers nothing. Adding a field that needs every product page may raise the cost.

Interactive result tables

Sort, filter, search, and shape the result before you export it.

• Show or hide columns and adjust their order.
• Find duplicates and choose a deduplication key.
• See source URLs and collection timestamps where available.
• Incomplete records and validation warnings are visible.
• A partial job is labeled partial. The exported selection is explicit.

An agent that builds the collector

For common shapes the agent starts from a tested, versioned recipe. For anything new it plans a collector, writes it, runs it in isolation, and validates the result before the pipeline becomes reusable.

• Recipes carry supported sources, output fields, pagination, deduplication, validation rules, and a last-verified date.
• A cheaper model handles known recipes. A stronger model is brought in for new compositions and recovery, inside explicit limits.
• Generated code runs isolated with resource limits. Scraped content is data, never instructions.
• The agent plans and collects. Application code enforces billing, limits, and storage.

Estimates and spending controls

You see an estimate or range with explicit assumptions before anything paid runs.

• Volume is marked known, capped, or uncertain.
• Run-level and account-level limits are enforced in application code.
• Budget is reserved per job so concurrent runs cannot spend the same allowance twice.
• Actual usage is reported after completion.

A hard price cap is backed by enforceable execution bounds, or by our decision to absorb an overrun. A stop button alone is not enough, because a provider may already have charged.

Exports and destinations

CSV in the first release. Google Docs reports and Airtable are planned. Google Sheets is a proposed addition.

• Export the full dataset or the filtered selection, explicitly.
• External destinations require you to connect them and pick where data goes.
• Repeated exports never duplicate or overwrite without a configured behavior.

Scheduled tasks

Save a validated workflow and run it on your frequency, in your timezone, within your limits.

• Fresh snapshot, append new records, or update matching records.
• Review previous runs, next execution time, and actual costs.
• Pause, resume, edit, or delete. Runs never overlap.
• Notifications on completion, meaningful change where supported, or failure.

Basic schedules are in the first release. Improved change tracking follows.

Hosted execution

We manage execution, job state, scheduling, stored results, and delivery. A job continues after you close the page.

• Retention periods, dataset limits, and supported volumes are defined before paid launch.
• Hosted pricing accounts for the resources a job uses.

Pages you can only see when signed in

Some data only appears after you log in: a dashboard, a members-only directory, a portal your company already pays for. Later, the agent will browse those pages the way you do, in a real browser session, using your own account.

• You sign in once, in a browser window we open for you. We never see or store your password.
• Only where the site’s terms allow it. The agent checks before it starts and says no when they do not.
• A higher tier or pay-as-you-go, because running a real browser costs more than reading a page.
• Everything else stays the same: columns you choose, an estimate you approve, a table with sources.

This is the one place where “anything online” currently stops. We are telling you now so you know it is on the list, not because it is available.

Downloadable workflow and code

Export a workflow package and run it in your own environment with your own provider accounts.

• Runnable code with pinned dependency versions.
• A configuration template and a credential template without secrets.
• A step-by-step how-to file, expected output fields, and known limitations.
• Estimated provider usage and the external costs that remain.

Running locally removes our hosting cost. It does not make provider APIs or model usage free.

Desktop application

A simpler way to configure and run downloaded workflows on your machine. A future distribution option, not a current feature.