L7Lesson 1

Wire Your First MCP

After this, you'll be able to install a real MCP server, connect it to your agent, and complete one actual task using it.

Before you start

Before diving in, complete Session Log vs Rules File so you have the session-log discipline to trace what each tool call did and write a rule when something goes wrong.

The idea

Here is the before and after: every session you spend 5 minutes copying your database schema into the chat so Claude can write correct queries. After wiring one MCP server, the agent pulls the schema itself. That 5 minutes drops to zero, every session, compounding indefinitely.

MCP (Model Context Protocol) is the standard that lets the model reach outside its context window and act on your real systems. You install an MCP server (a small program that runs locally), register it in your settings, and Claude gains a list of tools it can call. Each tool has a name, a description, and a parameter schema. Claude reads those descriptions and decides when to call the tool based on what you ask. You never have to say 'use the filesystem MCP.' It figures that out from the description.

For Claude Desktop users: open Settings, go to the Extensions or MCP tab, and add servers from the built-in list or by pasting a server URL. No terminal needed. For Claude Code users: add the server block to your .claude/settings.json, then run 'claude mcp list' to confirm it loaded.

Start with one server that removes a real copy-paste step. The filesystem MCP lets the agent read and write files in your project. The GitHub MCP lets it read PRs and check CI status. The Postgres MCP queries your dev database directly. Five minutes of setup, zero minutes of manual copying forever after.

Try it (20 min)

Watch out for

Installing three MCPs at once before you understand one. Start with one server and one task.
Registering the server but never verifying it loaded. Ask Claude 'what tools do you currently have access to?' to confirm.
Using production credentials for your first MCP. Always start with a read-only dev environment.
Expecting the agent to use the MCP automatically. Give it a task where the tool is clearly the right path.
Skipping the tool description. The model decides when to call tools based on descriptions. A vague description means the tool gets ignored.

Paste this into Claude:

I want to wire my first MCP server. Here is my current setup: [describe your stack and whether you use Claude Desktop or Claude Code]. Help me pick the best MCP to start with from this list: filesystem, Postgres, GitHub, Slack, or Figma. Once you recommend one, walk me through the exact steps to install it for my setup, register it in my settings, and verify it is connected. Then give me one real task I can complete using it right now.

What good looks like:

You chose one MCP server based on a real gap in your workflow, not because it sounded impressive
The server is installed, registered, and Claude acknowledges it in your session
You completed one real task using the MCP (not a test, an actual task you would have done manually)
You can name the tool the MCP exposed and describe what it does in one sentence
You verified the connection worked by seeing Claude call the tool and return real data

What a good response looks like:

Here is what a successful first MCP connection looks like for the GitHub MCP:

Claude's confirmation after connecting:
'I can see the following tools from the GitHub MCP server:
- list_pull_requests(owner, repo, state) — lists open/closed PRs for a repository
- get_pull_request(owner, repo, number) — fetches diff, description, and CI status for one PR
- create_issue(owner, repo, title, body) — opens a new issue

I do not have write access to push branches or merge PRs with these credentials.

For your first real task: I pulled up PR #47 in your repo. The CI check failed on the lint step. Here is the specific error from the workflow log: "ESLint: 3 errors in src/api/auth.ts (lines 14, 28, 31)." Want me to read those lines and suggest the fix?'

What you should see: Claude names the tools, describes what each one does, acknowledges what it cannot do, and immediately proposes a real task using live data from your repo.

When this breaks

Breaks when you wire many servers before mastering one because every additional schema loads into the system prompt and the model picks the wrong tool from a longer ambiguous menu
Breaks when production credentials are wired into the first server because any tool-call mistake during the learning phase touches real data instead of a sandbox
Breaks when registration is treated as success because a server that is registered but never invoked has zero compounding value and gives a false sense of capability

Claude can do it for you

Say to Claude: 'I have [MCP server name] connected. What tasks in my current project could you now do with this tool that you could not do before? Give me three concrete examples and then do the most valuable one.' Let it demonstrate its own new capability. For Claude Desktop: you can also just open a new conversation and ask 'what tools do you have?' after adding the server in Settings.

You can now

Demonstrate that the MCP server is connected by completing one real, previously manual task end-to-end and naming the specific tool the agent invoked.

Key takeaways

One working MCP that solves a real problem is worth more than five installed MCPs you never use. Wire it, use it once for real, then decide if it earns its place.

Wire one server before you wire many; verify it loaded by asking Claude what tools it has access to
Pick the MCP that removes a manual copy-paste step you do in every session, not the one that sounds impressive
Start in a read-only or dev environment so first-week mistakes don't touch production data
Tool descriptions are the trigger; the model picks the wrong tool when descriptions are vague or interchangeable

Go deeper

I want to wire my first MCP server. Here is my current setup: [describe your stack and whether you use Claude Desktop or Claude Code]. Help me pick the best MCP to start with from this list: filesystem, Postgres, GitHub, Slack, or Figma. Once you recommend one, walk me through the exact steps to install it for my setup, register it in my settings, and verify it is connected. Then give me one real task I can complete using it right now.

Here is what a successful first MCP connection looks like for the GitHub MCP: Claude's confirmation after connecting: 'I can see the following tools from the GitHub MCP server: - list_pull_requests(owner, repo, state) — lists open/closed PRs for a repository - get_pull_request(owner, repo, number) — fetches diff, description, and CI status for one PR - create_issue(owner, repo, title, body) — opens a new issue I do not have write access to push branches or merge PRs with these credentials. For your first real task: I pulled up PR #47 in your repo. The CI check failed on the lint step. Here is the specific error from the workflow log: "ESLint: 3 errors in src/api/auth.ts (lines 14, 28, 31)." Want me to read those lines and suggest the fix?' What you should see: Claude names the tools, describes what each one does, acknowledges what it cannot do, and immediately proposes a real task using live data from your repo.

Say to Claude: 'I have [MCP server name] connected. What tasks in my current project could you now do with this tool that you could not do before? Give me three concrete examples and then do the most valuable one.' Let it demonstrate its own new capability. For Claude Desktop: you can also just open a new conversation and ask 'what tools do you have?' after adding the server in Settings.