The problem this solves
Give the tool <company.com> and it returns 25 companies that look like it - by what they do, who they sell to, their size band and their stack - each with headcount, industry and a similarity score, as a JSON array your agent can loop over. That is the whole contract, and it is the one your own agent needs the moment a user says find me more like this one.
Building the equivalent yourself means stitching a company database to a similarity model to a firmographics source and keeping all three fresh. Through MCP the tool is already there, with a typed schema your agent framework can read, and it is the same similar-companies capability the Company Intelligence Agent uses in the console.
What the recipe has to cover is the plumbing your framework expects: the server URL and key, the tool's input and output schema so the model calls it correctly, a sample response so you can write the handler before you spend a credit, and the errors worth handling.
How the mission runs
- Connect your agent to the server. Your agent's MCP client points at the server URL with a Bearer key minted for this agent and scoped to company data reads. With the tools=all option on the URL the individual tools are listed and the similar-companies tool appears with its schema; without it the server offers mission_agent alone, which takes a whole objective in plain language.
- Read the tool schema. The Company Intelligence Agent prints the schema your framework will see: a required seed domain, an optional count that accepts 25, and optional filters such as a country, a headcount band or an industry to keep the lookalikes inside a market. The output schema describes an array of company objects and a block describing the resolved seed.
- Call it with a seed domain. The example call passes <company.com> and a count of 25. The seed is resolved first - firmographics, technographics and what the company does - and the lookalikes are ranked by how closely they match on those dimensions. The response is JSON: for each company, domain, name, headcount, industry, HQ country and a similarity score between 0 and 1, plus the seed profile so your agent can show what it matched against.
- Handle the sample response. With the sample response in hand your handler can be written before the first metered call: loop the array, drop rows below a similarity floor, and pass the survivors to whatever comes next - a people search for the buyer role, a fit score against your ICP definition, or a write to a list.
- Handle errors and cost. The recipe covers the cases: a seed the platform cannot resolve returns an empty array with a reason; a rate-limit response is retried after a pause; a credit-envelope response means stop and resume later. Each response carries the credits it metered, so your agent can report the cost of a lookup to its own user.
The prompt
This is the exact objective the agent receives. Swap the obvious placeholders for your own domain, segment or channel and run it as-is from the console, Slack, or the API.
What comes back
The MCP connection details for your agent, the similar-companies tool schema as your framework reads it, an example call with <company.com> as the seed, and a sample response of 25 lookalikes with domain, name, headcount, industry, HQ country and similarity score, plus the resolved seed profile. Your agent gains a find-more-like-this capability in one tool registration, with every call metered against the workspace and visible in the audit trail under the agent's key.
Make it yours
- Chain a fit score after the lookup so the 25 lookalikes come back ranked by your ICP definition rather than by similarity alone.
- Pass three seed domains - your best three customers - and merge the results, weighting companies that appear for more than one seed.
- Save the lookalikes to list <name> so they become a persistent, refreshable list the Prospecting Agent can add people to.
- Call the same tool through the REST API when your agent framework has no MCP client; the schema and the response are identical.
Frequently asked questions
What makes two companies similar?
What they do and who they sell to, their size band, their industry and the technologies detected on their properties, weighted together into a single score. The resolved seed profile in the response shows which of those the platform could establish, so a thin seed explains a loose match.
Can I restrict lookalikes to a country or size?
Yes, through the optional filters on the tool: a country, a headcount band, an industry. Filters apply before ranking, so you still get 25 rows when the market has them and a shorter array when it does not.
Is the result the same every time?
It is computed from live data, so the array moves as the underlying companies change, which is the point of calling it rather than caching a list. When your agent needs a fixed set, save the result to a list and read the list.
Which key should my agent use?
A key minted for this agent alone, scoped to company data reads. One key per surface means the agent's usage is metered under its own name, its rate limit is its own, and revoking it touches nothing else.