What MCP servers are, and why your agency should run one

The short version

An MCP server is a small program that sits between Claude and one of your tools (CRM, database, Notion, GitHub, etc.) and translates. You run it. Claude connects to it. Claude can now read and write to that tool on your behalf. The server defines exactly what Claude is allowed to do — no more, no less.

Think of it as a USB port. Claude is the computer. Each tool is a peripheral. MCP is the plug.

Why you should care

Without MCP, if you want Claude to interact with your CRM, someone has to build a custom integration. It takes a week. It breaks when the CRM changes. It's limited to whatever that developer decided to expose.

With MCP, you install (or build) an MCP server once. Claude — any Claude, Desktop, Code, API — can use it immediately. When Anthropic improves the model, your integration doesn't break.

For an agency, this means two things:

1. You can connect Claude to a client's systems in a day, not a month.
2. You can productise your tools: the same MCP server works for every client who runs that same stack.

What an MCP server actually does

At its simplest, an MCP server exposes tools and resources.

Tools are actions. "Create a contact in HubSpot." "Log a note on deal #4521." "Run this SQL query." The server defines each tool's inputs (what Claude has to provide) and outputs (what it gets back).

Resources are readable data. "All open deals." "This Notion page." The server describes the data; Claude requests it when relevant.

When you chat with Claude and mention "create a HubSpot contact for Sarah Chen," Claude looks at the MCP tools it has available, sees one that matches, and — with your permission — calls it. Sarah ends up in HubSpot. No copy-paste.

The ones that already exist

Anthropic maintains official MCP servers for common tools: Slack, GitHub, Google Drive, Postgres, Filesystem, Puppeteer. The community has built many more — Notion, Linear, Airtable, Sentry, Stripe, and a hundred others. You can install them in minutes.

For most agencies, the starter set is:

• Filesystem — Claude can read and write local files. Useful for drafting documents.
• Postgres — Claude can query your database. Useful for analytics.
• GitHub — Claude can read and comment on issues, create PRs. Useful for devs.
• Slack — Claude can read channel history and post. Useful for ops.

Three commands in a config file and Claude Desktop has them all.

When to build your own

Off-the-shelf MCP servers are fine for public tools. The moment you want Claude to interact with your systems — your custom CRM, your internal dashboard, your legacy SQL database — you need your own MCP server.

This is not hard. The official Anthropic SDK makes it a few hundred lines of code. The hard part is deciding what to expose.

Our rule: expose the minimum surface area that makes Claude useful. Every tool you expose is a door. Each door is a risk. Start with read-only tools. Add write tools only when you trust the pattern.

The security story

This is the part people overthink. MCP doesn't change your permissions model — it uses it. If Claude connects to your Postgres via MCP using a read-only role, Claude is read-only. If you give it write access, Claude can write. The MCP server is a seatbelt, not a new car.

What MCP does give you is visibility. Every tool call is logged. You can see exactly what Claude did and when. No more mystery about what the AI was up to.

Why this matters for small companies

For the first time, you can stand up a Claude integration with your internal systems in a day. Not a quarter. Not a consulting project. A day.

This changes what's worth building. Custom internal tools that used to cost 50,000 now cost 5,000. CRM integrations that took a month now take an afternoon. The economics of "we should automate this" shifted under our feet.

If your agency is still quoting custom integration work at 2024 prices, you're about to lose that bid.

What we build for clients

For clients on our Software pillar, the first thing we often ship is a small MCP server that connects Claude to their internal stack. The second thing is usually a dashboard that consumes it. The third thing is the thing they actually asked for, which is usually far smaller than they'd feared.

---

The official MCP docs are at modelcontextprotocol.io. Start there. Come back if you want us to build it for you.