AI chat
Describe the data you need: the chat finds an Actor, runs it, and summarizes the results. When it asks for approval, and how it handles prompt injection.
The AI chat finds Actors for you, runs them, and answers from their results. You describe what you need, for example "find flats under €1,500 in Berlin with a balcony", and the chat searches the Store, fills in the Actor's input, starts the run, and summarizes what it found. It can also manage your account: saved inputs, automations, runs, storages, webhooks, and more.
Open AI Chat in the side menu of the dashboard. The chat answers in the language of your first message.
What it can do
| You ask | The chat |
|---|---|
| Find data | Searches the Store, picks an Actor, runs it, and shows the results as a table with a short summary. |
| Run an Actor or saved input | Starts it right away with the input you described. |
| Read a page | Reads web pages for you with a crawler run. |
| Automate | Prepares an automation, such as "every morning, run my two job searches and send me a summary on Slack". |
| Change your account | Creates or edits saved inputs, automations, webhooks, API keys, notification settings, and more. |
| Share results | Makes a link to a dataset or key-value store that works without signing in, with the expiry you pick: 1 hour (the default), 1 day, 7 days, or 30 days. |
When no Actor in the Store does what you need, the chat can send the Scrape.ly team a request for one.
The chat can't change your password, sign-in or two-factor settings, billing and card details, or rotate your Default API key; use the dashboard for those.
Runs and approvals
A run you ask for starts right away, without a confirmation step. The chat shows an approval card first only when:
- the run is expected to cost more than $3, or you asked for a cost cap above $3. The card shows the expected cost and, when the run has one, the Maximum cost;
- the Actor needs full permissions and you haven't approved it yet;
- one message would start more than 100 runs, as a safety net against loops.
Chat runs have no default spending cap. A run gets a Maximum cost per run only when you ask for one, for example "run it, but spend at most $2". Runs are normal runs on your account: they're billed like any other run, show in Runs, and count toward your plan's concurrency and memory limits. When your run slots are full, the chat waits up to 2 minutes for a free one.
Changes to your account, such as creating an automation, editing a saved input, or making a share link, always show a confirmation card first. Nothing changes until you click Confirm.
Protection against prompt injection
Web pages, dataset rows, run logs, READMEs, and other tool results are written by third parties and can contain text that tries to steer the chat. The chat treats all of it as data, never as instructions:
- Once such content is in the conversation, a run of an Actor or saved input that only that content names doesn't start on its own. You get an approval card, whatever the cost.
- A run input or page read that would send your own data (account details, storage contents, run results) to a site that only tool results mention also waits for your approval.
- Share links to your storages leave the chat, in a run input or a page read, only with your approval, unless you typed the link yourself.
- What you clicked on a card is never read as something you wrote.
Costs and limits
- The chat's own AI model usage is billed to your account at the model's cost × 1.5.
- Runs the chat starts are billed like any run. See What a run costs.
- On the free plan, the chat answers 30 messages a day (UTC). See Account.
What's next
- Run Actors: input, run options, and limits
- Automations: schedules, filters, the AI step, and deliveries
- MCP server: use Scrape.ly from your own AI agent
