Claude Code MCP Server Setup: Stop Tab-Switching and Run Your CRM From the Terminal

You ship product from the terminal. Your business ops, meanwhile, are scattered across five to eight separate tools, and none of them talk to each other. Every lead lookup means a tab switch. Every call log means a detour through a CRM UI you never meant to open. By the third context switch of the morning, the focus you had at 9 a.m. is gone.

A Claude Code MCP server closes that gap. It exposes your CRM, lead data, and outreach as callable tools, right where you already work. Ask the agent to find a buyer and it searches. Tell it to log a call and the note lands in your CRM. The loop stays in one window.

This guide covers the full setup. You'll learn what a Claude Code MCP server does, how to wire one to your go-to-market stack, and how to keep an agent on a short leash while it works. Examples reference Zrevick, which speaks MCP natively. The steps apply to any business-grade server. By the end you'll have a working connection, a few tested workflows, and guardrails you actually trust.

What a Claude Code MCP server does, in plain terms

MCP is an open protocol for connecting LLM applications to outside data and tools (MCP specification). It shuttles JSON-RPC messages between three roles:

• Host: the LLM app. Here, that's Claude Code.
• Client: the connector inside the host.
• Server: the service exposing context and capabilities.

If you know the Language Server Protocol, the shape will feel familiar. The spec says MCP draws inspiration from LSP, which standardized language tooling across editors (MCP specification). MCP does the same for AI apps and everything else.

Here's the practical version for a founder who codes. A Claude Code MCP server is a program that hands Claude new tools. Without one, the model can read your repo and run shell commands. With a business server attached, it gains functions like search_leads or log_activity. Claude picks the tool, you approve the call, and the result lands in your CRM.

That answers a question most tutorials skip: what can a Claude Code MCP server actually do for go-to-market work? Whatever the server exposes. Find buyers, enrich records, draft sequences, move pipeline stages, publish content. The model becomes an operator, not just a coder.

Anatomy of a business MCP server: tools, resources, durable state

The spec defines three server primitives (MCP specification):

Tools are the verbs. Resources are the nouns, and prompts are the boilerplate.

A business-grade Claude Code MCP server needs a fourth quality the spec doesn't name: durable state. Chat context evaporates. Your CRM can't. When an agent enriches a lead, that write must outlive the session. Otherwise every run starts from zero and your data drifts.

This is where dev-tool servers and business servers diverge. Zrevick backs its MCP surface with durable data states and execution run histories. A lead enriched at 9 a.m. is still enriched next Tuesday, with a trace of what touched it.

Step-by-step: connect a Claude Code MCP server to your CRM and outreach stack

This is the part most people search for: connecting Claude Code to external business tools. The flow is the same for any server. Add it, check it, use it (Claude Code docs).

First, install Claude Code and authenticate.

1. Add the server

Run this in your terminal, not inside a claude session (Claude Code docs):

claude mcp add --transport http zrevick <your-zrevick-mcp-endpoint>

Adding a Claude Code MCP server takes one command. The parts:

• claude mcp add registers a server with Claude Code.
• --transport http means the server is hosted at a URL, not run locally.
• zrevick is a name you pick. Claude Code labels the server's tools with it.
• The last argument is the endpoint URL. Substitute your own.

Many business servers need a token. Servers that take a static token accept it at add time (Claude Code docs):

claude mcp add --transport http zrevick <endpoint> \
  --header "Authorization: Bearer <token>"

Others use environment variables; pass them with --env KEY=value. Check your provider's docs for which method it uses.

2. Verify the connection

claude mcp list

Watch the status column. ✔ Connected means ready. ! Needs authentication means you sign in via /mcp inside a session. ✘ Failed to connect means the server didn't respond; run claude mcp get zrevick for the error detail (Claude Code docs).

Inside a session, /mcp lists your servers, shows their tools, and handles authentication.

3. Pick a scope

Servers live in one of three scopes (Claude Code docs):

• local (default): private to you, this project only.
• user: private to you, active in all your projects.
• project: written to .mcp.json in the repo root, shared with your team.

For a solo founder, --scope user is a sensible default. You probably want your CRM everywhere, not stuck in one repo.

4. Use it

Start a session and hand the agent a real task:

Use the zrevick server: pull up the Acme Corp contact and summarize recent activity.

The first tool call asks permission. Approve it. Tool calls in the output carry the server name, so you can confirm which server answered.

5. Optional: commit .mcp.json as config-as-code

For teams, project scope doubles as version-controlled configuration (Claude Code docs):

{
  "mcpServers": {
    "zrevick": {
      "type": "http",
      "url": "<your-zrevick-mcp-endpoint>"
    }
  }
}

Claude Code reads .mcp.json at session start. The first time it sees a project-scoped server, it prompts for approval. That prompt exists so a cloned repo can't launch processes on your machine without consent (Claude Code docs). One caution: a bearer token inside .mcp.json can ride into version control. Keep tokens at user or local scope.

Quick troubleshooting: a slow first start can hit the default 30-second timeout; raise it with the MCP_TIMEOUT environment variable. A server that connects but lists no tools is usually missing a required environment variable (Claude Code docs).

Example workflows: enrich, draft, and log from the terminal

With a Claude Code MCP server connected, the day-to-day looks like this. The examples below are illustrative; exact tool names depend on the server you connect.

Enrich a lead

Find the VP of Engineering at Acme Corp. Enrich the record and add them to the Q4 outbound pipeline.

The agent searches, enriches, and writes. No tab-switching, no CSV uploads. In Zrevick, the enriched record lands in the Identity Graph, one profile shared across sales, outreach, and support.

Draft outreach

Draft a three-touch sequence for the Acme lead. Mention their Series A. Keep each email under 80 words. Save as draft, don't send.

The agent reads the lead's profile as context, drafts the sequence, and stores it. Note the closing instruction: draft, don't send. Writing and sending are different permissions. Keep them separate while you calibrate trust.

Log the call

Log my 2 p.m. Acme call: they need SOC 2 before signing. Next step is security docs on Friday. Move the deal to Verbal.

Post-call notes take seconds instead of a detour into the CRM UI. Better yet, the note feeds the shared company record. The next outreach draft and the support workflow both see it. Each interaction sharpens the next one.

Execution run histories: debug agent actions like code

Agents fail. When one does, you want what you'd expect from any flaky system: a trace.

Zrevick keeps execution run histories for automated workflows: durable records of what each run did. When outreach misfires or a record looks wrong, don't guess. Open the run history and read it.

Then debug like you would code:

1. Reproduce the failure from the run's inputs.
2. Find the first step that returned something unexpected.
3. Fix whatever the trace points to, whether that's the prompt, the tool config, or the data.
4. Re-run and compare the two histories.

Build this habit early. Treat your first scopes and prompts as starting points to test, not final settings. The run history is the feedback loop that makes each revision safe.

Security and scope: what an agent should and should not do

MCP is powerful, and the spec is blunt about the risk. Tools represent arbitrary code execution and must be treated with appropriate caution (MCP specification). The protocol can't enforce safety on its own. Implementers are told to build consent flows and access controls.

So how do you secure a Claude Code MCP server connection? Start with these rules:

1. Keep approval prompts on. Hosts must obtain explicit user consent before invoking any tool (MCP specification). Don't batch-approve everything on day one. Understand each tool before authorizing it, as the spec advises.

2. Separate drafting from sending. Let the agent write emails, sequences, and content. Approve sends yourself until the error rate justifies relaxing. Treat any threshold you pick as an experiment, not a fixed policy.

3. Scope credentials narrowly. Prefer a dedicated API token over your personal one where your provider offers it. Keep tokens out of committed files. Where available, use the OAuth browser sign-in flow instead; Claude Code supports hosted servers that authenticate this way (Claude Code docs).

4. Distrust tool descriptions. The spec treats tool behavior descriptions as untrusted unless they come from a trusted server (MCP specification). Connect only servers you control or have vetted.

5. Keep destructive actions out of scope. Deletions, payments, and irreversible writes shouldn't be agent-callable at first. Add them only after run histories show clean behavior.

6. Mind the data boundary. A GTM server holds your customers' contact data. Hosts must not transmit resource data elsewhere without user consent (MCP specification). Decide what your agent may read before you connect it.

The line to draw: agents execute, humans own outcomes. Approvals, scopes, and run histories keep that line visible.

The bigger picture

Most Claude Code integrations so far target the dev loop: issue trackers, browsers, error monitors, design handoffs (Claude Academy). Business operations have stayed manual by comparison. A business-grade Claude Code MCP server closes that gap.

You now know what the server exposes, how to add and verify it, and where to draw the security lines. Zrevick builds for exactly this. One MCP surface covers CRM, enrichment, outreach, content, and support, backed by a shared Identity Graph. Plans start around $34 per month billed annually, with unlimited workspaces, seats, and AI agents.

For a technical founder, the pitch is simple: your business stack becomes as scriptable as your product. Install Claude Code, run the add command with your Zrevick endpoint, and enrich your first lead without leaving the terminal.