Connecting Claude to live B2B data takes three steps: pick an MCP data server, authorize it in Claude's connector settings with OAuth or an API key, and define your ideal customer profile so results are scored for you. Setup takes about ten minutes, and from then on prospecting happens in the chat.
This playbook uses Signl as the worked example because we can show every step exactly. The shape is the same for any server in the category; if you are still choosing one, we compared the field in the best MCP servers for sales, and if MCP itself is new to you, start with what an MCP server for sales is.
Step 1: Create an account and choose your auth path
Get set up in the provider's dashboard (for Signl, that starts with a demo call). You then have two ways to connect, and the choice matters more than it looks:
- OAuth (recommended). In Claude, open Settings, then Connectors, then Add connector, and paste the server URL. Claude redirects you to the provider to approve access, and you never handle a secret. Access is revocable per client.
- API key. Generate a key in the dashboard and paste it into the client's configuration. Simpler for scripts and CLI use, but the key is a bearer secret: anyone holding it is you. A well-built provider shows the key once and stores only a hash; if yours can re-display a key in plaintext, that is a red flag worth acting on.
Use OAuth for Claude and ChatGPT. Keep API keys for automation you run yourself.
Step 2: Verify the connection with a throwaway query
Before doing real work, confirm the agent can see the tools. Ask Claude:
What Signl tools do you have access to, and what does each one cost?A healthy setup lists the tools with their billing behavior (search free, enrichment charged on delivery). If Claude says it has no such tools, the usual causes, in order: the connector was added but not authorized, you are in a different workspace or project than the one you connected, or the client needs a restart to refresh its tool list.
Step 3: Run your first real search
Now give it an actual brief. Be concrete about the market and the trigger:
Find UK logistics companies with 50 to 200 employees that
raised funding in the last 60 days and are hiring sales
roles. Show me the ten best matches with their signals.Two things to expect from a good server here. First, the answer should come back in seconds and reflect the live market, not a quarterly crawl. Second, searching should cost nothing: on Signl, search is free and unlimited on every plan, so you can iterate on the brief as many times as you want before spending a credit.
Step 4: Define your ICP so every answer improves
This is the step most people skip, and it is the highest-leverage one. In the dashboard, define your ideal customer profile: industries, headcount range, geographies, funding stages, and the signals you care about. Once saved, every search, weekly drop and standing query is scored against it automatically, with no extra prompting.
The difference is visible immediately: unscored results give you companies that match filters, scored results give you companies ranked by how much they look like your actual buyer.
Step 5: Enrich deliberately, and read the result codes
When the list looks right, ask for contacts:
Get me the top decision-maker at the three best matches.
For the first one, I want the VP of Sales specifically.On Signl this is the only step that costs credits, and only when a verified email is actually delivered; a miss returns a machine-readable reason instead of a charge. That structure matters beyond billing: when the agent knows the difference between "no verified email found" and "insufficient balance", it can decide to skip, retry elsewhere, or tell you to top up, instead of inventing an excuse.
The failure modes to expect in week one
Before any of these, it is worth reading the Signl MCP tool reference: most week-one confusion is a tool being asked for something it does not take.
- Tool confusion. Two connected servers with similar search tools make agents pick inconsistently. Keep one data server per workspace until you have a reason not to.
- Invented filters. Agents sometimes ask a server for filters it does not have, get an error, and silently drop the constraint. If a result set looks too broad, ask the agent which filters were actually applied.
- Stale-data trust. If your server answers from a crawl, treat any "recent" signal with suspicion and check the date on it. Live-query servers make this a non-issue.
- Session amnesia. A new conversation does not remember last session's shortlist. Pin the brief and the shortlist in a project, or re-paste them; do not assume continuity.
Ten minutes of setup, one saved ICP, and the loop from "who should we talk to" to a verified contact runs inside the chat you already use. The fastest way to know if that is worth it for your team is the free tier: point Claude at your real market and count what comes back.