How to Build an AI Sales Agent (and See How It Works)

An AI sales agent is a loop. A model reads a goal, calls tools to fetch data or act, checks the result and stops when a rule says so. Everything else is packaging.

This guide sits in the LeadOcean blog and builds the smallest useful version: an agent that sizes a market, pulls ten prospects and enriches one. You will see each call and its output.

TL;DR

• An agent is a model plus tools plus a stop rule. The data tool decides how good it is.
• This build takes about 30 minutes and spends about 10 of your 1,000 free records.
• Sizing a list is free. Only returned people cost a record.

Prerequisites

• A free LeadOcean account and API key, exported as $LEADOCEAN_API_KEY (sign up, no card).
• curl and jq for the raw calls.
• Claude Code, or any MCP client, for the agent version. LeadOcean's hosted MCP server is at https://api.leadocean.io/mcp and uses OAuth sign-in, so no key is pasted.

Step 1: Understand the loop

Every sales agent runs the same four moves. Observe the goal. Pick a tool. Read the result. Decide whether to continue.

goal: find 10 VPs of Sales at US companies with 51-200 staff
loop:
  1. count_leads(filters)        -> is the list big enough? (free)
  2. search_leads(filters, 10)   -> fetch rows (1 record each)
  3. get_lead(person_id)         -> enrich the best one (1 record)
  4. stop when 10 rows are held or the budget is spent

The model does the picking. The tools do the work. Your stop rule is what keeps it from spending records on a loop.

Step 2: Give the agent a data tool

Connect the MCP server so the model can call LeadOcean tools itself. In Claude Code the command is:

claude mcp add --transport http leadocean https://api.leadocean.io/mcp

Claude Code's MCP docs (September 2026) describe this --transport http form. Then run /mcp in Claude Code, pick leadocean and authenticate in the browser. The full setup is on the Claude Code MCP page. There are 13 tools. leadocean_count_leads, leadocean_count_companies, leadocean_list_filters, leadocean_list_enum_values, leadocean_get_account and the export status tools are free. Search, lead, phone and company lookups cost one record per person or company returned.

Step 3: Size the market before spending anything

Ask the free count first. An agent that skips this step burns records on empty or huge lists.

curl -s -X POST "https://api.leadocean.io/v1/people/search?count=true" \
  -H "x-api-key: $LEADOCEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "jobLevel": ["VP"],
    "jobFunction": ["Sales & Business Development"],
    "country": ["US"],
    "employeeRange": ["51-200"],
    "limit": 1
  }' | jq '.meta'

count=true goes in the query string, not the body. Enum values are exact strings, so resolve them with leadocean_list_enum_values instead of guessing.

Step 4: Fetch the rows

Now pull ten people. The body is flat JSON, filters are arrays and limit runs from 1 to 100.

curl -s -X POST https://api.leadocean.io/v1/people/search \
  -H "x-api-key: $LEADOCEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "jobLevel": ["VP"],
    "jobFunction": ["Sales & Business Development"],
    "country": ["US"],
    "employeeRange": ["51-200"],
    "limit": 10
  }' | jq '{count: (.data | length), next: .meta.nextCursor}'

That spends 10 records. To page further, pass cursor from meta.nextCursor. One search pages to 10,000 rows at most, then meta.depthCapped is true.

Step 5: Enrich one record and decide

Enrichment takes exactly one of linkedin_url, email, phone or person_id. There is no name-plus-domain lookup, so the agent searches first and enriches by person_id.

curl -s -X POST https://api.leadocean.io/v1/people/enrich \
  -H "x-api-key: $LEADOCEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"person_id": "PERSON_ID_FROM_STEP_4", "reveal_email": true}'

An enrichment that finds nobody returns a 404 and costs nothing. The response includes career history in resume_data. A good agent uses it: a VP who joined three months ago is a different message than one with six years in seat.

Step 6: Add the stop rule

This is the part tutorials skip. Without it, an agent retries until your records are gone.

Rules for the agent:
- Call leadocean_count_leads before any search. If the total is over 20,000, narrow the filters.
- Never fetch more than 10 people per run.
- Enrich at most 3 people per run.
- If a call returns 402 or 429, stop and report. Do not retry.

Put these in the system prompt or a CLAUDE.md file. A hard cap in your own code is safer than a polite instruction.

What you get

Step 3 is the free sizing call. We ran it on 2026-10-21 through leadocean_count_leads. Filters: VP, Sales & Business Development, US, 51-200 employees. Result:

{
  "total": 18567,
  "totalIsExact": true,
  "totalAll": 60456,
  "distinctCompanies": 12192,
  "credits": 0
}

total is the mailable audience: people whose best email is verified, catch_all_valid or catch_all. totalAll is the same filters with no email condition. The call cost zero records.

Step 4 returns people in the shape below (response shape from the API reference, placeholder values only):

{
  "success": true,
  "data": [
    {
      "profile_data": { "profile_full_name": "Jane Doe", "profile_headline": "VP Sales at Acme" },
      "contact_data": {
        "has_email": true,
        "email_status": "verified",
        "contact_current_experiences": [
          { "job_title": "VP Sales", "company_name": "Acme", "company_domain": "acme.com" }
        ]
      },
      "meta": { "source_ids": { "person_id": "123456789" } }
    }
  ],
  "meta": { "nextCursor": "CURSOR" }
}

The person_id for Step 5 is at meta.source_ids.person_id. Search rows carry email flags, not the address.

Troubleshooting

FAQ