How to connect your AI coding agent to your own tools with MCP

If you have spent time with an AI coding agent, you have probably hit the wall where it says something like "I can't see your database" or "paste the issue details here." The agent is smart, but it is boxed in — it only knows what is in your files and what you type. The Model Context Protocol (MCP) is how you knock down that wall and let the agent reach your real tools: a database, GitHub, a monitoring dashboard, a design file.

This guide explains what MCP is and walks through connecting your first server to Claude Code, with the same ideas applying to Cursor and Copilot.

What MCP actually is

MCP is an open standard — think of it as a universal adapter between AI agents and the systems you already use. Instead of every tool inventing its own way to talk to every agent, a tool ships one MCP server, and any MCP-aware agent (Claude, Cursor, Copilot) can connect to it. That is why you have seen "MCP" everywhere in 2026: it is the plumbing that lets agents do real work instead of just talking about it.

A good rule of thumb: connect a server whenever you catch yourself copying data into the chat from another tool. If you keep pasting error logs from Sentry or ticket details from Jira, that is a signal an MCP server would save you the trip.

The two kinds of servers

There are two flavors you will meet:

• Remote (HTTP) servers run in the cloud and you connect over the network — Notion, GitHub, and most SaaS tools work this way. This is the recommended default.
• Local (stdio) servers run as a process on your own machine — useful for things that need direct system access, like a local database driver.

Adding your first server

The core command is claude mcp add. Here is the basic shape for a remote HTTP server:

claude mcp add --transport http <name> <url>

A concrete example — connecting Notion so your agent can read and update pages:

claude mcp add --transport http notion https://mcp.notion.com/mcp

If the server needs authentication, pass a token as a header:

claude mcp add --transport http secure-api https://api.example.com/mcp \
  --header "Authorization: Bearer your-token"

For a local server, everything after -- is the command that starts it:

claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable \
  -- npx -y airtable-mcp-server

That -- matters: it separates Claude's own flags from the command that launches the server, so a --port meant for the server does not get parsed as a Claude option.

Once added, check what you have wired up:

claude mcp list        # see all servers and their status
claude mcp get notion  # details for one server

Inside a Claude Code session, the /mcp command shows live connection status and, for servers that use OAuth, lets you sign in.

Choose the right scope

When you add a server, decide who should see it with the --scope flag:

• local (the default) — just you, just this project.
• project — shared with your whole team through a checked-in .mcp.json file. Great for onboarding: a new teammate clones the repo and the agent already knows about the team's tools.
• user — available to you across every project on your machine.

For a team database or issue tracker, project scope is usually what you want. For a personal API key, keep it local or user so it never lands in the repo.

Do it safely

MCP is powerful precisely because it gives an agent real reach — so a little caution pays off:

• Trust the server before you connect it. A server that fetches external content can expose you to prompt-injection risk, where malicious text in the fetched data tries to steer your agent. Stick to official servers from vendors you recognize.
• Scope credentials tightly. Give the server a read-only or least-privilege key when the task only needs to read. This is exactly the principle Microsoft baked into its enterprise Dataverse plugin — the agent only ever sees what its granted role allows.
• Keep secrets out of shared config. Put tokens in environment variables or a local scope, never in a project-scoped .mcp.json that gets committed.
• Review what it can do. After connecting, open /mcp and skim the tools the server exposes so there are no surprises.

The payoff

Once a server is connected, the change in how you work is immediate. Instead of "here is the schema I pasted," you can ask: "Find the 10 most recent signups in our Postgres database and draft a feedback-session invite for each." The agent queries the database, reads the results, and drafts the emails — all in one flow, because it can finally reach the systems your work actually lives in.

Start with one server for the tool you paste from most. That single connection usually makes the value obvious faster than any amount of reading about it.