mcp-handbook

What Is an MCP Server?

An MCP server is a translator between an AI assistant and a business tool. It lets AI models like Claude, ChatGPT, or Gemini actually do things in your software — pull data from your CRM, enrich a lead, send a message, update a spreadsheet. Without an MCP server, the AI can only talk. With one, it can act.

What does MCP stand for?

Model Context Protocol. Anthropic created it in late 2024 as an open standard. Google, OpenAI, and Microsoft have all adopted it since. Think of it as USB-C for AI — one universal plug instead of a different cable for every device.

Before MCP, connecting AI to business tools meant building a custom integration for every combination. 10 AI models times 10 tools meant 100 integrations. MCP changes the math: each tool needs one server, each AI model needs one client. 10 + 10 = 20 instead of 100.

What can you actually do with one?

The pattern is always the same: you describe what you want in plain English, and the MCP server gives the AI the hands to go do it.

  • Sales: "Find all companies in my CRM that haven't been contacted in 30 days, enrich them with current employee count, and draft a follow-up email for each one." The MCP server connects to HubSpot, runs the query, pulls enrichment data, and generates the emails.
  • Marketing: "Pull last week's campaign performance from Google Analytics, compare it to the previous week, and create a summary with recommendations." The MCP server connects to GA4, pulls the data, and produces an analysis.
  • Operations: "Find every deal in our pipeline stuck in the same stage for more than 14 days and flag them in Slack." The MCP server connects to Salesforce and Slack, runs the logic, and posts the alerts.
  • Engineering: "Check which pull requests have been open for more than a week and summarize the blocking issues." The MCP server connects to GitHub, queries open PRs, and returns the summary.

How it works (the simple version)

Three pieces:

  1. The AI model (Claude, ChatGPT, Gemini) — the brain. It understands your request and figures out what needs to happen.
  2. The MCP server — the translator. It sits between the AI and your tools, converting intentions into actual API calls.
  3. Your business tools (CRM, database, email, Slack) — the systems where your data lives and where work gets done.

When you ask Claude to "add this contact to my CRM":

  1. Claude understands you want to create a CRM record
  2. Claude calls the MCP server's create_contact tool
  3. The MCP server translates that into a HubSpot API call
  4. HubSpot creates the contact and confirms
  5. Claude tells you it's done

You never see steps 2-4. You just ask, and it happens.

Who builds MCP servers?

  • Software vendors build them for their own products. HubSpot, GitHub, Slack, Datadog, and hundreds of others now ship official MCP servers.
  • Companies build internal MCP servers for proprietary systems — internal databases, custom tools, private workflows.
  • Independent builders create MCP servers for tools that don't have official ones yet. The Clay MCP Server is an example — 73 tools built by a practitioner because the vendor hadn't shipped one.

Is it safe?

MCP includes permission controls — read vs. write access, scoped data access, audit trails. The key questions to ask about any MCP server:

  • What data can the AI access?
  • What can it change vs. only read?
  • Is there a log of what happened?
  • Who built it and who maintains it?

Open-source servers are auditable. Vendor-provided servers follow the vendor's security model. The protocol itself doesn't bypass your existing access controls — it works within them.

Where to go next

If you want to see what an MCP server looks like for a specific tool, start with one you already use:

FAQ

What does MCP stand for?

Model Context Protocol. An open standard created by Anthropic that lets AI models connect to external tools through a universal interface.

Do I need to be technical to use one?

No. Using an MCP server is as simple as talking to an AI assistant. Building one requires programming knowledge, but using one is just natural language.

Is MCP only for Claude?

No. Google, OpenAI, Microsoft, and most major AI platforms now support it. Build once, use everywhere.

How much does an MCP server cost?

The protocol is free and open source. Costs come from the tools you connect to and the AI model you use. Many MCP servers are also free and open source.

What's the difference between an MCP server and a regular API?

A regular API integration is point-to-point — one app talks to one service. An MCP server is a universal adapter: any compatible AI model can use any MCP server without custom integration code.