Skip to content

Find leads: agents that find new businesses to sell to

Last updated November 7, 2026

On this page

Agents can work only with the customers a company already has, unless they can look outside. Find leads lets an agent do the looking a salesperson does by hand: search Google Places and the web for businesses that fit ("salons in Pune with a 4.5 rating or better"), read a business's own website, drop the ones the company already has, score how well each fits, and put a batch in front of the right sales lead. Nothing found becomes a lead until a person approves it, and nothing approved is contacted until a person turns outreach on.

This page covers the tools and what the kernel checks on every call, the Lead Finder, Company → Find leads (providers, monthly limits, Google Places, reading websites, batches, where approved leads go, territories), the Prospects page and its batch cards, a found lead's source line, outreach, what is kept, and the API. The steps use the demo company (ok demo serve): Ada is the admin, Sam leads Sales, Alice and Ben are reps. What is not possible yet is listed at the end.

Words used on this page#

  • Prospect: a business an agent found, held apart from the CRM until a person decides. It keeps where it was found (Google Places, a web page), when, and why it fits.
  • Batch: up to 50 prospects found together, for one approver to review on one card.
  • Territory: cities and localities, the person who approves their batches, and the reps who own the leads made from them.
  • Already known: a business that is already a lead, a contact, a prospect, or a record in a connected system (by place id, phone or website). It is dropped before anything is paid for and never offered again.
  • Possible duplicate: a business whose name and locality are close to one already known ("Lotus Salon Baner" and "Lotus Hair & Beauty, Baner"). It is shown to the reviewer and never merged.

The tools#

Five tools are built into the kernel rather than reached through MCP, so the kernel can enforce what a page or a provider may do:

Tool What it does Cost
web.search Searches the web through Brave Search or Tavily, with the company's key Billed per search
web.read Reads one public web page the run already read the address of Free
places.search Finds places on Google Places by what they are, where, and a minimum rating; returns place ids only Free
places.names Reads the names and areas of place ids a search returned, to catch a business that may already be known before paying for details Billed at a cheaper rate
places.details Reads one place's details (the fields the company chose) Billed per place

Every result is page text or data from outside the company: a run that read one works in review mode, and the text reaches the model as quoted data, never as instructions ("ignore your instructions" on a page is read as words on a page). The kernel, not the model, finds the phones and emails on a page and says whose they are: a phone from Google Places or from the business's own site (its contact section, header or footer) is the business's; a mobile beside a person's name, or a number on someone else's page, is a person's, and is kept restricted (pii.prospect). A model not cleared for restricted data never sees it.

What the kernel refuses, with the sentence the run reads:

Refused Sentence
Reading an address the run did not read (a made-up URL) "That address was not read in this run"
A link on a row that no page the run read gave and no other source holds (a search result alone is not enough) "That link was not read in this run: read its page with web.read first; a search result alone is not enough"
A phone or email on a row the run did not read "That phone or email was not read in this run"
A page robots.txt asks readers to leave alone "{site} does not allow reading this page (robots.txt)"
A site whose robots.txt did not answer "That site could not be checked: its robots.txt did not answer"
A blocked site (LinkedIn, Google search and Maps pages, and by default Justdial, IndiaMART, Facebook and Instagram) "{site} is blocked for reading"
A call past the provider's monthly limit "{provider} reached this month's limit of {amount}"
A search carrying a person's details (a lead's personal phone or email) "A search may not carry {class} data"

A redirect is followed at most three times, and every hop is checked again: a link that passes through a blocked site, or to an address inside the company's network, is refused at that hop. Pages are read with GET only, without cookies or sign-in, at most one page a site every two seconds (a second agent reading the same site waits), under the user agent OrchKernel (+<your contact page>). A run acting for nobody may not use the tools, except the Lead Finder on a schedule an admin set.

The Lead Finder#

The Sales app's Lead Finder holds the four tools and the prospects.write_batch tool that puts a batch up for review. Any other agent gets the tools only when an admin adds them to its toolset. Ask it the way you would ask a colleague, in a conversation:

Find salons in Pune with 4.5+ rating that are not in our CRM

It recalls the company's profile (categories, minimum rating, cities), reads what the CRM already holds for the city, searches Google Places for ids (free), reads the names of those ids at a cheaper price, reads details only for ids that are not already known or possibly known (billed), reads the businesses' websites, scores each from 0 to 100 with a one-line note, and writes a batch. Its answer says how many it found, how many were already known, how many went in the batch and what it cost. It also runs on a repeat, and as a goal's "new prospects" tactic (see Goals).

It never writes a phone, email or link the run did not read, never guesses an email from a name, never logs in to a site, and never contacts anyone.

Company → Find leads#

Admins open Find leads under Company in the sidebar (/settings/find-leads). One Save at the end saves everything but the territories, which have their own.

Search services. Add Google Places, Brave Search or Tavily. Each uses a connection holding your company's own key (Apps → Connections, made for that service). For each, say the most it may cost in a month, in dollars: every billed call counts toward it, and at the limit agents stop calling that service until the next month. $0 stops every billed call; the page says so when you type it. Beside each service is this month so far ("This month: $1.80 of $3.00 · 300 calls"); Google Places' figures are counted at list price, before Google's free monthly usage, so the meter can show more than Google bills. With both Brave and Tavily set up, web searches use the first listed.

Google Places. Tick what agents read about each new place (name, address, area and city, rating, number of reviews, phone, website, opening hours). The line under the boxes says what one place costs to read with those fields, and Google's name for it ("Place Details Enterprise"). A place already known is never read.

Keep everything from Google Places is on by default: prospects keep what Google gave of the fields ticked above (name, area, address, rating, reviews, phone, website and hours); a lead keeps the name, with the address, hours, phone, website and place in its notes. A field left unticked is not read, so there is nothing to keep. A company that turned on the Sales app before version 4 keeps address and hours once it updates the app; until then the page says so and the rest works as before. Beside it, always: "Google's terms allow keeping only the place id; Google may suspend your Maps key." Off, prospects and leads keep only the place id (and a lead its business name), neither address nor hours, and the name, phone and rating are read again from Google, billed, when someone opens them. What a run read from Google goes when the run ends.

Where to look. Countries (two letters, such as IN), cities and kinds of business.

Reading websites. Your company's contact page, named to every site agents read. Read even where robots.txt says not to: only for sites that agreed to it. Also allow lifts a default block (Justdial, say); Never read adds your own. LinkedIn and Google's search and Maps pages are never read, whatever the lists say.

Batches. How many places one batch holds (at most 50), how long a rejected place is kept with its reason (90 days unless you change it), and who approves a city no territory covers. With nobody chosen, agents cannot put up a batch for such a city: "No territory covers {place} and no default approver is set; an admin names one under Settings → Find leads".

Where approved leads go. Leads here, in the CRM app (the default), or a connected system: choose the connector, the tool that creates a lead there (create_lead), what the record is called there (lead), and how its fields map to ours, one a line (business_name=name; address and hours are ours too, and go only where you map them). The rep who owns a lead is matched to a person there by email.

Territories#

Under the settings, Territories lists each territory: its name, the person who approves its batches, its cities and (optionally) localities, and the reps who own its leads. Leads go to the reps in turn; a rep who has left is skipped, and the page says so. A place goes to the first territory that lists its locality, else the first that lists its city, else to the approver for uncovered cities.

Import reads a connected system's territories and team members (dingg-ats' list_territories and list_team_members), matching people by email: a territory's manager approves (a role such as Manager, Sales Manager or Head of Sales, not a Key Account Manager), its other members are reps. The imported list is shown to check, and saved only with Save territories. The import runs as the agent that holds those two tools and acts for you (give the Lead Finder the connector's toolset first).

Try it: as Ada, add Google Places with a limit of 3, add a territory "Pune" with Sam approving and Alice and Ben as reps, and save both.

The Prospects page#

Prospects (in the Apps section once the Sales app is on, /prospects) shows the batches you may see: those you approve, those of your territory, and every batch for an admin. To review shows the batches with rows still waiting; All batches shows the rest too.

A batch's card says where the rows came from ("Pune: Google Places"), who found them and for whom, when, the counts ("12 found · 12 already known, 12 skipped before paying · cost $0.18 at list price") and who approves it. A batch whose rows were all known already reads "Nothing new", and nobody is called to it. Each row shows the business's name (linked to where it was found), its area, rating and reviews, phone, website, address and hours, its fit score and note, and its state. A possible duplicate is marked "Maybe the same business as …: check before approving". A person's own contact (a mobile beside an owner's name) is marked as a person's.

Its approver (or an admin):

  • Approve a row, after choosing the rep who will own it when the territory has more than one. The row becomes a lead, owned by that rep, with where it was found.
  • Reject a row, saying why ("already a customer", "closed"). The reason is kept; rejected rows are removed after the days the settings say.
  • Approve all or Reject all the rows still waiting.
  • Turn on outreach for the approved rows (below).

Two approvers clicking at once never make two leads from one row. Everyone else reads the card without the buttons. With "Keep everything" off, a row shows Show details from Google (billed to the agent that found it, at the price it shows) instead of its name and phone. It reads when you press it, not by itself, so opening the page costs nothing.

A found lead's source line#

A lead the Lead Finder made shows, on its page and in the Brain's record drawer, where it came from: "Found on Google Places, 4 Nov, by Lead Finder", with a Source link, then whether outreach is on.

Outreach waits for a person#

A lead made from a prospect starts with outreach off, and changing its email or adding a contact to it does not change that. No outreach task starts for it, and any send to it or its contacts, from any path (an SDR ask, a goal's tactic, a company trigger, an outside agent), is refused before anyone is asked to approve it: "Outreach is off for this lead; a person must turn it on". Approving a goal's plan does not turn it on. Its owner, its batch's approver or an admin turns it on, for one lead on its source line (Turn on outreach) or for a batch's approved rows on the card, which starts its outreach tasks. Then the company's consent rules and limits apply as to any send (see Contacts and consent and Limits): under the India preset, a WhatsApp message to a number found this way needs a recorded opt-in.

What is kept#

  • Prospects with where, when and by whom they were found, until decided; rejected ones with their reason, for the retention days.
  • For each found lead, its source, owner and whether outreach is on.
  • Each provider's spend and calls per month, in numbers only (never a query or a result).
  • The contacts each run read, until the run ends, so a run parked on an approval still finds them after a restart. A call outside a run keeps its own for 30 days, except what Google gave with "Keep everything" off.
  • A keyed hash of each address a batch row held, so outreach stays off for it whatever "Keep everything" says, and nothing of Google's is kept for this. A rejected row's are dropped at once. A row nobody decided stops holding its addresses after the retention days, and it never stops a send to an address the company already holds on a lead or a contact.

What happens to search results#

Web search results last only as long as the run that asked for them, from either provider, so changing provider changes nothing about what is held.

  • While the run lives. The run reads them and carries on with them. A run parked waiting for a person keeps them until it ends, because it needs them to go on, across a restart too. An outside agent's work session, should one hold these tools, keeps them until the session ends.
  • When it ends. However it ends (it finishes, fails, is stopped or the server stops under it), every title, snippet and link goes. No later run reuses them. Its step reads, for example, "Searched for salons in Pune: 12 results. Search results aren't kept after the run." What stays is the query, the provider, the number of results, the cost and the time.
  • A prospect row keeps what a page said, read with web.read, or what another source gave (Google Places, Apollo, a connector), never what a search result alone listed. A search result only says where to look.

A company wanting to keep results longer would need its own right from the provider, under its own agreement, which this product doesn't use. The rule is the same for Brave Search and Tavily.

What this cannot reach:

  • An agent's answer is in its own words, and those words may repeat what it saw. A plan cannot copy a search result into an answer, a message, a record or a memory by reference, but it can write the words out.
  • Backups hold the data until their own retention removes them, and the state file's free space holds a deleted row's bytes until SQLite reuses it.

API#

Every route is for people; an agent is refused with 403 people_only ("Finding leads is set up and reviewed by people; agents use the Find leads tools.").

Route What
GET /api/settings/find-leads The settings, their defaults, each provider's month (spend, calls, cap and what they are counted at), the price of a place with the chosen fields, the connections and connectors to choose from (admins)
PUT /api/settings/find-leads Save the settings (admins); the destination is { "to": "native" } or { "to": "connector", "connector", "create_tool", "kind", "fields" }
GET /api/settings/find-leads/quote ?mask=displayName,rating: what one place costs to read with those fields (admins)
GET /api/territories The territories (admins)
PUT /api/territories Save them (admins)
POST /api/territories/import { "connector": "dingg-ats" }: territories read from a connected system, to check and save (admins)
GET /api/prospects { batches, live_per_call }: the batches you may see, each with can_decide, and what one live read of a row costs
POST /api/prospects/batches/:id/decide { "decisions": [{ "prospect", "decision": "approve" | "reject", "owner", "reason" }] } (a decision without prospect is every row still waiting): the card after it (its approver or an admin)
GET /api/prospects/:id/live A row's place read live from Google, billed
POST /api/leads/outreach { "refs": ["leads:<id>"], "on": true }: { changed, refused }
GET /api/leads/source ?ref=leads:<id>: { source, can_turn_on }: where a found lead came from, or null, and whether you may turn its outreach on
GET /api/leads/live ?ref=leads:<id>: a found lead's place read live, billed

Errors: 403 denied (a member on the settings or the territories, or deciding a batch they do not approve), 404 not_found, 422 invalid with problems (every problem with the settings at once: "a batch holds 1 to 50 rows", "a rejected row says why").

Not possible yet#

  • Searching LinkedIn, logging in to sites or solving CAPTCHAs: never, by design.
  • Guessing an email from a person's name: never, by design.
  • Other places providers than Google Places.