Lead Research with MCP: A Step-by-Step Playbook

Lead research means turning a target profile into a short list of real companies and people worth contacting. A sales rep or SDR usually does it by hand, tab by tab, across a data tool, company websites and the CRM. With MCP, one AI assistant can query a prospecting database, read company websites and check your CRM in the same conversation.

This playbook uses three documented server types: Apollo.io for people and company data, a web search or scraping server for context, and your CRM to avoid duplicates. It ends before outreach, on purpose.

What you need

• Apollo.io MCP server. Apollo hosts it at https://mcp.apollo.io/mcp, with OAuth sign-in. Apollo's MCP product page says "There's no separate MCP subscription or additional cost," and lists native connectors in Claude, ChatGPT and Perplexity. See our Apollo.io MCP server profile.
• A web research server. Two documented options: the Exa MCP server (web_search_exa, web_fetch_exa, agent_run) and the Firecrawl MCP server (firecrawl_search, firecrawl_scrape, firecrawl_agent).
• Your CRM's MCP server. For HubSpot, see the HubSpot MCP server profile. Salesforce and Pipedrive also publish MCP servers.
• An AI client that supports remote Streamable HTTP servers. Apollo's README lists VS Code with GitHub Copilot, Cursor, Claude Code, Claude Desktop and Claude Cowork.
• A written ideal customer profile (ICP). Industry, size, region, buyer titles and disqualifiers. The assistant applies it; you own it.

The workflow

1. Turn the ICP into filters. Ask the assistant to restate your ICP as search filters and show them to you before running anything.
2. Search companies. apollo_mixed_companies_search searches organizations by industry, size, location and keywords, per Apollo's README. Keep the first list to a size you would actually review.
3. Check the CRM for overlap. Use your CRM's search tool, such as HubSpot's search_crm_objects, to mark companies that are already customers, open deals or recently lost. Drop or tag them.
4. Read each company's own pages. Use web_fetch_exa or firecrawl_scrape on the company's site for what it sells and who it serves. web_search_exa or firecrawl_search can find recent news. Firecrawl's README says search results carry excerpts and you scrape only the URLs you need.
5. Score against the ICP. Ask for a table: company, fit reasons, red flags, source URL for each claim. Reject anything without a source.
6. Find people at qualified companies. apollo_mixed_people_api_search searches by title, seniority, location and company attributes.
7. Enrich only the shortlist. apollo_people_match and apollo_organizations_enrich (or their bulk versions) return contact details. Apollo's README says "some actions (such as enrichment) consume Apollo credits."
8. Save approved leads. After review, create records with apollo_contacts_create or your CRM's write tool. In Claude, HubSpot caps creates and updates at 10 records at a time.
9. Stop before outreach. Sequence enrollment is a separate, human step (see Guardrails).

For bigger research jobs, Exa's MCP docs describe agent_run for "research that takes more than one search," and Firecrawl documents firecrawl_agent for multi-source research that returns structured data. Both are usage-billed; Exa says agent_run needs OAuth or an API key.

Example prompts

• "Turn this ICP into Apollo company search filters and show me the filters. Do not run the search yet."
• "Run apollo_mixed_companies_search with those filters, limit to 25. Then check each company in HubSpot with search_crm_objects and mark any existing records."
• "For the 10 companies with no CRM match, use web_fetch_exa on their homepage and pricing page. Summarize what they sell and who they sell to, with the URL for each fact."
• "Find VP or Director of Operations at the five best-fit companies using apollo_mixed_people_api_search. Do not enrich yet; tell me how many people you found."
• "Enrich these 6 people with apollo_people_match and draft a table for my review. Do not create contacts."

Guardrails

• Gate every credit spend. Search first, enrich second, and only after a person approves the shortlist. Apollo's README says its skills "warn and request confirmation before credit-consuming actions."
• Keep sequences manual. Apollo warns that apollo_emailer_campaigns_add_contact_ids "may start outbound depending on sequence settings." Apollo tells users to verify "the sequence name, sending account, and enrollment volume before confirming." In this playbook, a rep does that step.
• Follow privacy and anti-spam law. Apollo's README tells users to follow "applicable privacy, marketing, anti-spam, and data-protection laws." Enriched emails and phone numbers are personal data.
• Treat web pages as untrusted. A scraped page can contain text aimed at the AI. Read MCP prompt injection and keep write tools on approval while web servers are connected. The MCP Security Center has more.
• Use your own login. Apollo says the server acts with your Apollo permissions and plan limits. Do not share one account across a team.
• Never paste API keys into chat. Firecrawl's README says never to put a key in the server URL or an agent chat.

What to measure

• Share of AI-suggested companies a rep accepts after review.
• Share of accepted leads that turn into meetings, compared with your manually researched leads over the same period.
• Apollo credits used per accepted lead.
• Duplicate records created by mistake (target zero).
• Bounce rate on enriched emails once a person sends to them.
• Claims in research summaries that failed a spot check against the source URL.

Next step

Pick one segment, run the workflow for 20 companies and compare the result with a rep's manual list. More task guides are in the MCP playbooks hub. If you want this wired into your CRM with the right approvals, see our MCP development service.

Frequently asked questions