MCPs: Connecting Claude to Your Stack
12 min readMCPs give Claude live access to your tools · Notion, Gmail, Google Calendar, GitHub, your CRM. Here's what they are, how to install your first one, and the core stack to connect.
What is an MCP?
MCP stands for Model Context Protocol · the open standard for connecting AI agents to external tools and services.
Without MCPs, Claude can only read and write files on your local machine. With MCPs, Claude has live access to your external stack:
- Read and update your Notion pages and databases
- Search your Gmail inbox and draft replies
- Check your Google Calendar and create events
- Push commits and manage repos in GitHub
- Read and update deals and contacts in your CRM
The difference is significant. A Claude session without MCPs is a very smart assistant sitting in a room with no phone and no internet. MCPs give it a line to the outside world. This is what turns a skill from "drafts a follow-up" into "drafts the follow-up, sends it from your inbox, and logs it to the deal."
How to install an MCP
The general pattern:
claude mcp add [name] [transport flags] [server-url-or-command]Each MCP has its own install command. Here's the Notion MCP to try right now:
claude mcp add --transport http notion https://mcp.notion.com/mcpFirst run: Claude opens your browser to authenticate with your Notion account. After that, it's connected permanently.
Test it by asking Claude:
Read my [page name] page in Notion and give me a summaryWatching Claude read a real page from your own workspace is usually the moment it clicks. It stops feeling like a tool and starts feeling like something that actually knows your world.
The core stack to connect
You don't need all of these on day one. Install the one for the tool you use most, then add the rest as you need them.
| MCP | Best for | Install |
|---|---|---|
| Notion | Knowledge management, project context, your brain | claude mcp add --transport http notion https://mcp.notion.com/mcp |
| Gmail | Email drafting, inbox search, follow-up automation | Add via the Gmail MCP server URL and authenticate in-browser |
| Google Calendar | Scheduling, daily briefings, availability checks | Add via the Calendar MCP server URL and authenticate in-browser |
| GitHub | Code management, repo creation, commit history | claude mcp add --transport http github https://api.githubcopilot.com/mcp |
| CRM (HubSpot / Attio) | Pipeline tracking, contact logging, deal updates | Add via your CRM's MCP server URL and authenticate in-browser |
Each MCP authenticates the same way · run the add command, Claude opens your browser, you grant access, it's connected for good. To see what's connected at any time, run claude mcp list.
A connector being installed is not the same as it being active. With remote MCPs (the ones you add through the connectors panel), Claude treats each chat like a fresh instance · if the connector isn't toggled on for that specific conversation, Claude will tell you it can't browse / can't reach the tool. First thing to check when an MCP "isn't working": open the connectors panel and make sure it's green. For local config-based MCPs, the equivalent gotcha is a malformed JSON config · usually a stray trailing comma. Paste it into jsonlint.com when in doubt.
Two worked examples
The pattern is the same every time · the only thing that changes is what you're plugging in. Here are two that turn Claude into something most people pay for.
Give Claude a real browser · TinyFish
Out of the box, Claude can't check a live web page, pull local business data, or operate a site. TinyFish is an MCP that hands it an actual browser it can navigate, click, and pull structured data from. Add it as a custom connector with this remote URL:
https://agent.tinyfish.ai/mcpGrab a free key at tinyfish.ai first, then: Claude → + → Add Connector → Manage Connectors → + Add Custom Connector, name it TinyFish, paste the URL above, click Connect, and complete the sign-in. Toggle it on in the chat, then try:
Find the top 20 pizza restaurants in Austin, Texas. For each, pull
name, Google rating, review count, website, and phone. Return as a table.That's outbound research · the kind of thing agencies pay Apollo or Clay for · running on any city or niche.
Push cold email into your sending stack · Smartlead
Claude writes great cold email but can't land it in the inbox · that's deliverability infrastructure (warmup, domain health, inbox rotation). Smartlead handles that layer. There's no first-party Smartlead MCP yet, but you don't need one: Smartlead has a REST API, and Claude Code can call it directly with your key.
Get the key from Smartlead → Settings → Integrations → API Keys → Generate API Key, and keep it in your project's .env as SMARTLEAD_API_KEY (never paste a key into a chat). Then describe the outcome and let Claude draft and create the campaign:
Write a 4-step cold email sequence for [ICP] selling [offer], hook off [pain].
Then create it in Smartlead as a NEW campaign via their REST API (key in .env as
SMARTLEAD_API_KEY), and leave it PAUSED so I can review before anything sends.Claude drafts the copy, calls the Smartlead API to create the paused campaign, and reports the campaign ID. Prefer to keep it manual? Drop the "via their REST API" line and just paste the drafted sequence into Smartlead's campaign builder yourself.
Make "leave it PAUSED so I can review before anything sends" a permanent part of the instruction · you review both the sequence and the lead list before a single email goes out.
MCPs + skills = real workflows
The real power is combining them. Example:
- You finish a client call
- You paste your raw notes into Claude and run a meeting-notes skill
- It outputs a structured summary, action items, and a follow-up email draft
- With the Gmail MCP active, Claude sends the follow-up directly from your inbox
- With the Notion MCP active, Claude logs the summary to the client's project
That's one command replacing twenty minutes of manual work. This is the pattern every AI employee on the SkillTree map is built around · a skill for the workflow, MCPs for the reach into your real tools.
Run a real first session
With Claude Code installed, your CLAUDE.md set, and at least one MCP connected, run an actual job end to end:
cd /your-project-folder
claudeThen orient Claude and hand it something specific:
I'm working on [project]. Today I want to [specific task].
Here's the relevant context: [paste any notes or details]Vague prompts get vague output. Be specific · "write 3 cold email variations to [prospect] at [company]; their pain is [pain]; my angle is [angle]; max 5 sentences each, one CTA per email" beats "help me with outreach" every time.
Before you finish, run /diff to see every file Claude touched, and /compact once the session gets long. If something goes wrong, screenshot the error and paste it back in, or /clear and retry with a tighter prompt. Errors are part of the process · the goal isn't to avoid them, it's to get fast at fixing them.
"The greatest mistake you can make in the early stages is to be impatient, to demand progress when the world is asking you to absorb."
Robert Greene, Mastery
That's the foundation. Claude Code is installed, configured, and connected to your stack. Next module: build your brain · the second brain every skill reads from.