How to Automate a Salesforce Lead Workflow with Claude Code (2026)

The reliable pattern is a script that Claude Code writes and you run. LeadOcean supplies the people, and Salesforce's sObject Collections upsert writes them as Leads, keyed on an external ID so a rerun never duplicates.

Keep the chat for design and the script for volume. This guide builds one workflow: a list of US sales operations people, landed in Salesforce as Leads.

The reason to script it is control. A chat session that writes to a CRM row by row is slow, hard to review and hard to rerun. A script with a dry run and an external ID is easy to check, safe to repeat and cheap to hand to a colleague.

TL;DR

• It takes about 45 minutes the first time and about 10 for each new list.
• Sizing the list is free. The 200-person test spends 200 LeadOcean records, which is 20% of the 1,000 free ones.
• Salesforce upserts up to 200 records per request, so 200 Leads is one write call (Salesforce REST API guide, September 2026).

Prerequisites

• A LeadOcean API key from app.leadocean.io, stored as $LEADOCEAN_API_KEY. Free has no card and 5 requests a second.
• A Salesforce sandbox, not production, with an access token and instance URL in $SF_ACCESS_TOKEN and $SF_INSTANCE_URL.
• A custom text field on Lead named LeadOcean_Key__c, marked External ID and Unique.
• Claude Code, curl and Python 3 with requests.

export LEADOCEAN_API_KEY="paste-your-key-here"
export SF_INSTANCE_URL="https://your-domain.my.salesforce.com"
export SF_ACCESS_TOKEN="paste-your-token-here"
claude mcp add --transport http leadocean https://api.leadocean.io/mcp

The last line adds the LeadOcean MCP server. It signs in with OAuth, so no key is pasted into the chat. sf org display --json prints the instance URL as result.instanceUrl. Current Salesforce CLI versions redact the access token in that output (salesforcecli/plugin-org source, October 2026), so take the token from your org's OAuth flow.

Step 1: Size the list before you spend anything

Ask Claude Code to count first. leadocean_count_leads is free however often you call it. Counting first tells you whether the list is 300 people or 300,000 before a single record is spent.

Use leadocean_count_leads with title "Sales Operations" and "RevOps", country US.
Tell me total, totalAll and distinctCompanies.

Claude Code sends this input.

{
  "title": ["Sales Operations", "RevOps"],
  "country": ["US"]
}

The same count over REST is a free search with count=true as a query parameter and limit=1.

curl -s -H "x-api-key: $LEADOCEAN_API_KEY" \
  "https://api.leadocean.io/v1/people/search?count=true&limit=1&title=Sales%20Operations,RevOps&country=US"

Title is a keyword match, so read the facets before you trust it. Narrow with jobLevel or seniority if the industry mix looks wrong.

Step 2: Export the people as a CSV

One export request returns a file you own. That beats paging search results through a chat window.

curl -X POST https://api.leadocean.io/v1/exports \
  -H "x-api-key: $LEADOCEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"name":"sfdc test 200","filters":{"title":["Sales Operations","RevOps"],"country":["US"],"emailStatus":["verified","catch_all_valid"],"emailType":["work"]},"limit":200,"columns":["profile_first_name","profile_last_name","current_job_title","current_company_name","current_company_domain","profile_url","email_1_address","email_1_status"]}'

Poll the export, then download it. The pollAfter field says how many seconds to wait.

curl -H "x-api-key: $LEADOCEAN_API_KEY" https://api.leadocean.io/v1/exports/EXPORT_ID
curl -H "x-api-key: $LEADOCEAN_API_KEY" -o people.csv https://api.leadocean.io/v1/exports/EXPORT_ID/download

An export takes up to 50,000 rows a request and 500,000 a day. Bulk CSV export is not on Free, so on Free run the same filters through search and enrich instead. The Exports page in app.leadocean.io builds the same file with a filter builder and shows the record price first.

Step 3: Give Claude Code a spec, not a vague task

Write the rules into the prompt. A script that cannot touch production beats a careful reviewer.

Write upsert_leads.py. It reads people.csv and upserts Salesforce Leads.

Rules:
- Skip rows where email_1_status is not verified or catch_all_valid.
- Map: profile_first_name to FirstName, profile_last_name to LastName,
  current_job_title to Title, current_company_name to Company,
  current_company_domain to Website, email_1_address to Email,
  lowercased email_1_address to LeadOcean_Key__c.
- LastName and Company are required on Lead. Skip rows missing either and count them.
- PATCH /services/data/v62.0/composite/sobjects/Lead/LeadOcean_Key__c in chunks of 200.
- Add LeadSource "LeadOcean".
- Default to a dry run that prints the first 3 payloads and sends nothing.
  Send only with --send.
- Read SF_INSTANCE_URL and SF_ACCESS_TOKEN from the environment. Never print them.

Replace v62.0 with your org's current API version. Review the diff before you run anything. Check that the field map matches your org, that the external ID field is the one in the URL, and that nothing prints a token.

Step 4: The upsert script

This is the shape Claude Code should produce. The endpoint, the 200 limit and the external ID in the path come from Salesforce's documentation (September 2026).

import csv, json, os, sys, requests

OK = {"verified", "catch_all_valid"}
URL = (os.environ["SF_INSTANCE_URL"]
       + "/services/data/v62.0/composite/sobjects/Lead/LeadOcean_Key__c")
HEAD = {"Authorization": "Bearer " + os.environ["SF_ACCESS_TOKEN"],
        "Content-Type": "application/json"}

records, skipped = [], 0
for r in csv.DictReader(open("people.csv")):
    if r["email_1_status"] not in OK or not r["profile_last_name"] or not r["current_company_name"]:
        skipped += 1
        continue
    records.append({
        "attributes": {"type": "Lead"},
        "FirstName": r["profile_first_name"],
        "LastName": r["profile_last_name"],
        "Title": r["current_job_title"],
        "Company": r["current_company_name"],
        "Website": r["current_company_domain"],
        "Email": r["email_1_address"],
        "LeadSource": "LeadOcean",
        "LeadOcean_Key__c": r["email_1_address"].lower(),
    })

print(f"{len(records)} to upsert, {skipped} skipped")
if "--send" not in sys.argv:
    print(json.dumps(records[:3], indent=2))
    sys.exit(0)

for i in range(0, len(records), 200):
    body = {"allOrNone": False, "records": records[i:i + 200]}
    res = requests.patch(URL, headers=HEAD, json=body)
    res.raise_for_status()
    results = res.json()
    failed = [x for x in results if not x["success"]]
    print(f"chunk {i // 200}: {len(results) - len(failed)} ok, {len(failed)} failed")
    for x in failed[:5]:
        print(x["errors"])

allOrNone: False lets good rows land when one row fails. Salesforce returns results in request order, so a failure maps back to a CSV line. Set it to True if you would rather roll back a whole chunk than clean up partial writes.

Step 5: Dry run, then send

Run it without --send first and read the three payloads. Then run it against the sandbox.

python3 upsert_leads.py
python3 upsert_leads.py --send
python3 upsert_leads.py --send   # second run updates, creates nothing new

The second send is the test that matters, because it proves the script is safe to schedule. Because LeadOcean_Key__c is the external ID, the rerun updates the same 200 Leads and creates none. Count Leads where LeadSource = 'LeadOcean' in the sandbox to confirm.

What you get

Step 1 on 2026-10-21, using the free leadocean_count_leads tool, titles "Sales Operations" and "RevOps", country US.

{
  "total": 34246,
  "totalIsExact": true,
  "totalAll": 81003,
  "distinctCompanies": 13034,
  "credits": 0
}

total counts people with an address worth sending to (verified, catch_all_valid or catch_all). totalAll is the same filters with no email condition. The largest industry in the facets was Software Development at 4,911 people. The call spent 0 records.

A person row from the API has this shape. Response shape from the API reference, placeholder values.

{
  "success": true,
  "data": [
    {
      "profile_data": {},
      "contact_data": { "has_email": true, "email_status": "verified" },
      "resume_data": {},
      "meta": {}
    }
  ],
  "meta": { "count": 1, "limit": 1 }
}

The export CSV is flat, with columns such as profile_first_name, current_job_title and email_1_address. A row looks like Jane,Doe,Director of Sales Operations,Acme,acme.com,verified, with placeholder values.

Troubleshooting

FAQ