Search base
An agent told "find where houses are sold in Madrid, and houses up to €450,000 with a plot of 2,000 m² or more" answers two questions in turn: where to search and what's there. For the first it builds a source registry, for the second a results database. Both live in the agent's notebook, and the easiest place to look at them is a separate dashboard section — "Search base" in the navigation on the left.
The section has five tabs: "Sources", "Results", "Automations", "Publications" and "Search quality". The base belongs to one agent: the one selected in the list.
How to start
Tell the builder:
Find all sources in Madrid where houses are sold, and houses matching our parameters
It will take the "Search agent" template, name an approximate price and wait for your "yes". The search runs as a long task: it's the most expensive work on the platform. An estimate (not a measurement) — 500–2,000 credits per pass, and a base of a hundred sources is 15–40 passes and several hours.
You don't need to set the topic and geography in advance: "Collect sites and channels about meditation across Spain" or "find packaging suppliers in Poland" — same agent, same tabs.
Sources
The three numbers at the top — found, active and on topic — are counted from the cards, not from the agent's words:
- active — sources whose latest publication was in the last 90 days;
- on topic — sources with relevance of 70 or more. The agent gives a score only with an explanation: the platform doesn't record a number without a word on why.
One site — one source. https://www.idealista.com/, idealista.com/en/ and es.idealista.com are the same card. For platforms with many sources (Telegram, Facebook, YouTube), a source is a channel or group, not the platform itself.
Access status is set by the platform, not by the agent: it opens the page itself and records the outcome and the check date.
| Status | What it means |
|---|---|
| readable | the page opened, the agent can read from it |
| blocked by robots.txt | the site asks not to read this address automatically — and the agent doesn't |
| login required | the page asks you to sign in, pay or confirm you're not a robot |
| not responding | the site didn't open or replied with an error |
| API required | the platform offers no open pages without signing in |
| not checked | the platform hasn't opened this source yet |
Select sources and press "Check again" — up to 25 at a time; it costs no credits. "Turn off" keeps the source in the registry, but the agent no longer visits it, and only you can turn it back on.
Your own list. If you already have sources — the "Import" button: links as text or a CSV or XLSX spreadsheet (the "name", "country", "language", "type" columns become source fields; CSV from Excel in Windows-1251 encoding is read too). The same in words: "here's my list of sites: …".
What the agent doesn't read
Without official access, the agent sees only what's open to everyone without signing in: sites, directories, forums, RSS feeds, public Telegram channels in web view. It doesn't read closed Telegram groups (invite-only).
It doesn't read or bypass Facebook, Instagram, TikTok, YouTube, WeChat and Weibo. Such sources stay in the registry with the "API required" status and an explanation: for the agent to work with them, connect official access with your own action — a key or token for that platform.
The platform doesn't bypass "I'm not a robot" checks, login pages or paid access either: the source gets the "login required" status, and there's no second attempt by other means. The agent doesn't promise "all groups" or "the whole internet".
Results
Each result is a card with normalised fields: price as a number and currency, floor area and plot in square metres (even if the page said "0.2 ha"), address, district, phone, date and a link to the page it was taken from.
Filter. "Add condition" — field, operation and value: price at most 450000, plot at least 2000. You can write the value the way you'd write a price: "€450,000", "0.2 ha" — the platform normalises it the same way as the card field. The filter lives in the page address: after a reload the table is the same, and you can send a colleague a link to it. The database compares across all cards — not the agent across the ones it remembers.
Under each row it says why it's there:
matches: price ≤ 450000 · not confirmed: plot ≥ 2000
"Not confirmed" means the card has no such field: the agent didn't find the plot area on the page. Such a card stays in the results below the confirmed ones, because "we don't know" isn't the same as "doesn't fit". The "Confirmed only" checkbox hides them.
A criterion that isn't in the cards at all — "under 20 minutes to the metro", "legally clean" — the platform doesn't confirm and doesn't pretend it did.
Search by meaning ("income property by the sea" finds "apartments, seafront") works only where a meaning-search model is turned on on the platform. Where there isn't one, the section says so: the search goes by words and fields. Field conditions always work.
The same property on several platforms — one row. The platform recognises it by street and house number, coordinates or name — and merges them itself only when everything that can be compared matches: price within a few percent, same area, same address. Two flats in one building, identical new-build listings, two offices of one company stay as separate rows marked "possibly the same" — you decide. Click the name: the card shows all platforms with the price on each. If the platform merged different things — "Unmerge", and it won't merge that pair again. Pairs it isn't sure about are marked "possible duplicate": deciding "same" or "different" is up to you.
Where a value comes from. In the card, each field has one of three marks:
- confirmed — the agent quoted a fragment of the page, and the value is in it;
- matches across sources — the same value appears on another platform;
- unconfirmed — the value was recorded without a fragment.
The platform doesn't store pages: you can check the fragment itself via the link.
Rating. "Fits" and "No" in each row are your rating. It doesn't disappear when the agent updates the card, and search precision is calculated from it.
Export. CSV and XLSX — the file gets exactly what the filter selected, not just the page shown.
To a CRM. Select results and press "Send to webhook": the platform shows the request body and sends it to your address (https only). If the address redirects elsewhere, the platform won't follow and will tell you: give the final address. There are no ready-made CRM connections — almost every CRM and every automation service accepts a webhook.
Automations
Set up conditions and press "Save search" — or say "save this search and run it every morning". The platform runs a saved search over the collected base itself: on a button press, hourly, daily or weekly. The tab shows the last run, how many match and how many are new since last time; "Show new" opens exactly those results. If something new appears on schedule, you get a notification.
A saved search costs nothing: it only looks at what's already collected. New cards are brought into the base by the agent — with a long task or a schedule — and that does cost credits.
Publications
The platform has no ready-made connections to other listing platforms. There are two ways.
By hand. The agent prepares text for the platform and language, tells you where to paste it and what to check. You post it yourself and press "Published" — with a link, if there is one.
Through your action. If the platform offers official access, connect it with an action. Then the agent puts the publication up for confirmation: it goes out only after your "Confirm and publish". Before that, the platform checks three things itself:
- whether this same text has already been published on this platform — a repeat isn't posted;
- whether the platform is on the "Do not contact" stop list;
- whether the limit is exhausted: no more than 20 publications a day and no more than once an hour per platform.
The "Publish via action without confirmation" switch removes only your "yes"; the three checks remain. And it applies only where nothing from someone else stood between you and the text: the "Publish" button in the dashboard under text you can see, and a publication you created yourself. Anything the agent composed or asked to post after reading someone else's page, an email or collected cards still waits for your confirmation — the card says why. A publication whose action parameters don't contain its text also waits: it's the parameters that go to the platform.
The publication status ("published", "awaiting confirmation") is set by the platform — from your decision, from the action's response or from the "Published" button; the agent can't write it. The agent doesn't rewrite a publication that's awaiting a decision or already posted: for a different text — a new card.
The log shows for each publication: where, in which language, what text, status, link, and if it failed — what the platform replied.
The platform doesn't do mass mailings.
Search quality
| Metric | What's counted |
|---|---|
| Coverage | how many sources from your reference list the agent found on its own. The list is added in "Sources" → "Import" marked "reference" |
| Precision | how many of the results you rated are "fits" |
| Freshness | how soon after a listing is published it gets into the base — where the card has a date |
| Merged duplicates | how many cards are the same property on another platform |
| Traceability | how many results have a link to the source page |
| Cost of a useful result | the agent's credits divided by results rated "fits" |
Where there's nothing to count from — you haven't rated anything yet, there's no reference list — it says exactly that, not "0%".
People's data
The agent keeps only contacts shown openly on pages, and for each one — the source (the page where it appears) and the basis (why it's here). The platform doesn't record a contact without them.
- A contact the agent hasn't touched for 180 days is removed by the platform itself. The period is changed in the "Search quality" tab or in words: "keep contacts for just a month".
- "Erase all contacts from source" removes contact cards from that site, and the phone and contact from results taken from it. In words: "erase the contacts from this site".
- Collected contacts are included in the workspace data export and disappear together with the agent.
How much fits
The agent's notebook holds 50,000 cards — sources, results, contacts and the publication log together; on the Pro plan — 150,000, on Business — 500,000. How much is used is shown in "Search quality". When there's no room, the agent hears it as a refusal with advice, and you see it as a "notebook full" problem.
The agent runs up to three long tasks at once, each up to 300 passes since the last "continue".
In words
| You say | What happens |
|---|---|
| "How many sources did you find? What's the search quality?" | registry numbers and metrics |
| "Here's my list of sites: …" | the list goes into the registry |
| "This is a reference list, compare against it" | coverage appears |
| "Check which sources can be read" | the platform checks access |
| "Turn off this source" | the agent no longer visits it |
| "Show houses up to 450,000 with a plot of 2,000 or more" | results by conditions |
| "Save this search and run it every morning" | a saved search with a schedule |
| "This one fits, this one doesn't" | ratings |
| "This is the same house" / "these are different" | duplicates |
| "Send the matching ones to the CRM" | webhook — after your consent |
| "I published it, here's the link" | an entry in the publication log |
| "Keep contacts for just a month" | retention period |
| "Erase the contacts from this site" | erasure — after your consent |