Playbooks / Recipe
Agent-Driven Account Research: Claude Pulls Company, People and Emails
Connect the MCP server once, then research an account by asking for it. Claude picks the endpoints, runs them, and hands back the execution_log receipt for each call. Only the email and phone waterfall calls are miss-free; the enrichment and search calls bill on a 2xx.
Pattern: One prompt → enrich_company → linkedin_company_employees_search → email_finder → email_verifier, chosen and sequenced by the assistant, run against your key and your credit pool.
In short: RichAPI runs a hosted MCP server at https://mcp.richapi.ai/mcp, authenticated with an x-api-key header, that exposes the same endpoints as the REST API at https://api.richapi.ai/api/v1/. An MCP client such as Claude Desktop, Claude Code or Cursor can therefore research an account end to end from a natural-language prompt: enrich_company (which takes a LinkedIn company page URL, not a domain), linkedin_company_employees_search or people_search, enrich_profile (also a LinkedIn URL), email_finder and email_verifier. MCP and REST share one API key and one prepaid credit pool. Every RichAPI tool is a read-only lookup; none send email, post content or write to a CRM. On the waterfall endpoints email_finder, email_verifier and phone_finder, a call that finds nothing returns success:false with billed:false and costs zero credits; all other endpoints bill on a successful 2xx response including an empty result. Per-result endpoints such as linkedin_company_employees_search bill per row returned. New accounts get 100 free credits with no credit card.
Last updated November 4, 2026
Credit math
One account, done properly: enrich_company 1 credit per call + linkedin_company_employees_search 0.1 credits per result × 5 results = 0.5 + enrich_profile 1 credit per call × 2 + email_finder 5 credits per call × 2 + email_verifier 2 credits per call × the ones that hit. Twelve accounts before a Monday pipeline review lands in the low hundreds of credits, and the finder misses inside that cost zero.
ROI math
Compare against the hour a rep spends with fifteen browser tabs open per account, and against the per-seat sales-intelligence licence that rep would otherwise need. Credits are pooled, not per-seat. See /pricing.
Owned internally by
The person doing the research. No engineer required after setup, which is the point.
Best for
AEs and founders prepping a small number of named accounts, and anyone who would rather type a sentence than build a script for a job they do twice a week.
- 01
Create a key at app.richapi.ai/signup.
100 free credits, no card.
- 02
Add the MCP server to your client config: the block above, or the per-client instructions at /integrations/claude. The host is mcp.richapi.ai and auth is an x-api-key header.
- 03
Restart the client and check the tool list.
The tool names map one-to-one onto the REST endpoints, so anything in the API is available in the chat.
- 04
Give the assistant a budget in the prompt.
'Stop at five people' and 'do not page further' are the difference between a 12-credit answer and a 200-credit one.
- 05
Ask for the cost column.
Every waterfall call returns its own billing line; making the assistant surface it turns an opaque session into a receipt you can read.
Research acme.com for a first meeting on Thursday.
1. Enrich the company from its LinkedIn company page URL: headcount,
industry, specialties.
2. Find up to five people there with "revenue operations" or "sales
operations" in the title. Stop at five; do not page further.
3. For the two most senior, find a work email and verify it.
4. Give me a table: name, title, email, verification status, and the
credits each call cost. Where a finder call missed, say which
providers were tried rather than leaving the row blank.
Agent-driven account research: one prompt, four endpoints, a receipt for each
Account research is a sequence of lookups that every rep does slightly differently and nobody writes down. Find the company. Work out who owns the problem. Get their email. Check it isn't dead. It takes an hour per account with a browser, or twenty lines of script you then have to maintain.
Through MCP it's a paragraph. The assistant picks the endpoints, fills the arguments from your prose, runs them against your key, and shows you what each call returned and what it cost.
The chain the assistant actually runs
Four steps, in this order, because each one produces the input for the next:
enrich_company— 1 credit per call. Headcount, industry, specialties. Takes a LinkedIn company page URL, not a domain. If you hand the assistant a bare domain it will needfind_website_by_company_nameor a search step first, and that costs a call. Give it the LinkedIn URL when you have it.linkedin_company_employees_searchorpeople_search— per-result, 0.1 credits per result. Title-filtered. This is where an unbounded prompt gets expensive: 200 rows returned is 200 rows billed.enrich_profile— 1 credit per call, also a LinkedIn URL, for the two or three people you actually care about.email_finderat 5 credits per call, thenemail_verifierat 2. Waterfall. A finder call that comes back with nothing returnssuccess: false,billed: false, and costs zero.
Steps 1 through 3 bill on any 2xx, including a thin one. Step 4 doesn't bill on a miss. That asymmetry is worth telling the assistant about, because it changes where caution is expensive: exploring with search calls costs money, retrying a finder does not.
The receipt is what makes this usable
The reason to do research in a chat window is speed. The reason not to is that you cannot see what happened. An assistant that reports "I couldn't find an email for Dana" is indistinguishable from an assistant that never made the call.
Every waterfall call returns an execution_log, an array of attempts with per-attempt status, plus a billed field. Ask for it in the prompt and the failure becomes legible: four sources tried, three returned no data, one timed out, nothing charged. Now you know whether to retry with the LinkedIn URL instead of name-plus-domain, or to accept the person isn't findable.
We don't publish the list of sources, because it changes as we add and drop them. The per-call log is the honest version of that answer.
The failure mode: an agent with no budget
The thing that goes wrong is not a wrong answer. It's a loop. You ask for "everyone in revenue at these twelve accounts", the assistant calls a per-result search endpoint with a generous limit, gets 300 rows per account, and enriches them because you said "everyone". Nothing errored. You just paid for 3,600 rows you will never read.
Three guards, all of them prompt-level:
- Cap the result count explicitly. "Stop at five." "Do not page past the first page." Per-result endpoints bill per row returned, so the limit is the budget.
- Name the endpoint when you care. "Use the company employee search, not a general people search" stops the assistant picking the broader tool.
- Ask for the cost column. An assistant that has to report credits per call behaves noticeably better than one that doesn't.
Start on the 100 free credits. If a prompt blows through those, it would have blown through a funded balance too, and you found out for nothing.
What it can't do to you
Every tool in the catalog is a read-only lookup. Nothing sends an email, posts anything, or writes to your CRM. The worst outcome of a badly phrased prompt is spent credits, not a message going out to a customer with an unfilled merge field in it.
There is also no CRM-native app. The assistant hands you a table; getting it into Salesforce is your integration.
Account research FAQ
Does this work in Cursor and other MCP clients, not just Claude? Yes. It's a standard remote MCP server. Client config formats differ; Claude carries the one we test.
Can the assistant spend my whole balance? It can spend credits, which is why every result carries a cost line and why the free tier exists. Cap result counts in the prompt.
Why does enrich_company reject my domain? Because it takes a LinkedIn company page URL. This trips up nearly everyone once. See company enrichment.
Is the MCP data different from the REST data? No. Same endpoints, same key, same credit pool, same responses. Prove a workflow in chat, then automate the version that worked against people search.
Do I need a credit card? No. 100 free credits on signup, enough to research one account properly.
Where to go after account research
The use cases hub has the rest of the plays. Claude for setup detail, list build and verify when the same chain has to run over 500 rows instead of 12, and waterfall email then verify for the billing model on the last two legs. Tiers on pricing.
100 free credits, no card. Connect the server and research one real account before your next call.
FAQ