Cédric Rittié

← Back to blog
7 min
claude codemcpautomation

MCP servers: connecting Claude to your tools

The MCP protocol turns Claude Code into a central hub. It reads your analytics, searches your contacts, posts to Slack, all without leaving the conversation.

Phase 3 · Automate · Article 1 of 3

The problem

You use Claude Code. You have your Skills, your CLAUDE.md, your conventions. Claude knows your project and works fast.

But it's deaf and blind to the rest of your world.

You want it to check the analytics from your latest campaign. It can't, PostHog is in another tab. You want it to look up a contact in your CRM. Impossible, it has no access to HubSpot. You want it to post a summary in Slack. It doesn't even know Slack exists.

Every time, you're the intermediary. You open PostHog, copy the numbers, paste them into the conversation. You open HubSpot, export a CSV, feed it to Claude. You write the Slack message yourself because Claude can't send it.

The AI does the analysis. You do the copy-paste between tools.

MCP servers remove that intermediary. They give Claude direct access to your tools. It reads, writes, and acts, without you leaving the conversation.

This article follows Phase 2 (Customize). Skills and CLAUDE.md are in place. Here, we connect Claude to the outside world.

MCP, a standard for all your tools

The most accurate analogy: MCP is a USB-C port for AI.

Before USB-C, every peripheral had its own cable. One for the printer, one for the camera, one for the hard drive. If you wanted to connect a new device, you needed a new driver, a new cable, a new setup.

MCP does the same thing for digital tools. Instead of building a custom connection between Claude and each service (one for Slack, one for Google Sheets, one for PostHog), MCP defines a single standard. One universal "cable."

Without MCP

You open PostHog. You copy the numbers.

You paste into Claude. "Analyze this data."

Claude responds. You copy the response.

You open Slack. You paste. You send.

4 tabs. 6 copy-pastes. 5 minutes.

With MCP

"Check this week's pageviews on PostHog and send a summary to #marketing on Slack."

Claude reads PostHog. Writes the summary. Posts to Slack.

1 sentence. 0 copy-pastes. 30 seconds.

How it works (no jargon)

An MCP server is a small program running on your machine that acts as a translator between Claude and an external service.

Claude Code Central hub
PostHog
Slack
Google Sheets
Gmail
HubSpot
Notion
Figma
Calendar

Each "node" is an MCP server. Install one, Claude gains a new ability. Install five, Claude accesses five services. The servers don't know about each other. They all talk to Claude through the same protocol. Claude orchestrates.

You don't write code to install an MCP server. You edit a configuration file and restart Claude Code. It takes 2 minutes.

Three concrete scenarios

No theory. Three real situations, three different roles, three MCP server combinations.

The weekly report that writes itself
Marketing

Every Monday, you spend 45 minutes opening PostHog, exporting numbers to a Google Sheet, writing a summary, and sending it to Slack.

1Claude reads this week's analytics via the PostHog MCP server
2It updates the tracking Google Sheet via the Google Sheets MCP server
3It writes a summary and posts it to #marketing via the Slack MCP server
Result: 45 minutes to 1 sentence. Every Monday.
The competitive audit in 10 minutes
Product

You want to compare your landing page to 3 competitors. Normally: open each site, take notes, check pricing, verify G2 reviews.

1Claude scrapes all 3 sites via the Firecrawl MCP server
2It searches for recent reviews via the Web Search MCP server
3It produces a comparison table and pushes it to a Google Sheet
Result: a structured comparison instead of 15 open tabs.
Research that feeds the newsletter
Communications

Every week, you read 50 articles, keep 5, write a summary for your newsletter. The rest disappears.

1Claude searches for the latest articles on your topics via Web Search
2It reads the full content of each article via Firecrawl
3It scores, filters and writes the synthesis using your research Skill
Result: 2 hours of reading compressed into 10 minutes of filtered results.

The pattern is always the same: Claude reads from one service, processes the information, and acts in another service. You stay in the conversation.

MCP servers to know

Not an exhaustive list. The servers that cover 80% of a product, marketing, or communications professional's needs.

PH
PostHog
Analytics
27+ tools. Pageviews, funnels, feature flags, errors. Claude reads your dashboards and answers questions in natural language.
"How many unique visitors this week on /blog?"
Sl
Slack
Communication
Read channels, post messages, summarize threads. Claude turns Slack into a searchable knowledge base.
"Summarize what was discussed in #product this week."
GS
Google Sheets
Data
Read, write, and analyze spreadsheets. The bridge between your raw data and Claude.
"Add this month's metrics to the April tab."
Gm
Gmail
Communication
Read emails, create drafts, search your inbox. Claude doesn't replace you, it prepares.
"Draft a reply to Pierre's email about the Q2 brief."
GC
Google Calendar
Organization
View events, find free slots, create meetings. Claude manages your calendar.
"Find a 30-minute slot with Marie this week."
No
Notion
Knowledge
Semantic search across your Notion workspace. Claude digs through your notes, docs, and wikis.
"Find the latest pricing decision in our wiki."
Fi
Figma
Design
Dev Mode access. Claude reads a Figma design's structure (hierarchy, tokens, styles) and generates faithful code.
"Generate the React component from this Figma design."
FC
Firecrawl
Web
Scrape any website. Claude reads the full content of a page, not just a search summary.
"Read Notion's pricing page and compare with ours."
HS
HubSpot
CRM
Deals, contacts, tickets. Claude queries your CRM in natural language.
"Which deals close this week and what's the total value?"
GH
GitHub
Project
Issues, PRs, reviews. Claude manages your backlog, creates issues, reviews PRs.
"Create an issue for the redirect bug Marie reported."
At
Atlassian (Jira)
Project
Jira issues, Confluence pages, search. Claude navigates your backlog and docs without opening Jira.
"Which tickets are blocked in the current sprint?"
Br
Brave Search
Web
Real-time web search. Claude accesses news, latest articles, and trends.
"What happened this week in the European UCaaS market?"

You don't need to install all of them. Start with 2-3 servers matching your daily tools. Add more when a concrete need arises.

How to install an MCP server

No code. A configuration file and a command.

Claude Code stores its MCP server list in a JSON file. You add the server you want, restart Claude Code, and it's active.

Add the PostHog MCP server.
Updating MCP configuration...
The PostHog MCP server is configured. You can now ask me questions about your analytics directly. Try "How many unique visitors this week?"

Most official MCP servers install in one command. For cloud servers (PostHog, Slack, Gmail), you typically need an API key or OAuth authorization. Claude guides you through the process.

What to check before installing:

QuestionWhy it matters
Is the server official or community-built?Official servers (Anthropic, the service provider) are maintained and secure. Community ones, check the GitHub repo.
What permissions does the server request?Read-only vs read/write. A read-only PostHog server is risk-free. A Slack server with write access can post on your behalf.
Does data pass through a third party?Most MCP servers run locally. Your data stays on your machine.

MCP vs Skills: two different things

Skills and MCP servers complement each other but do different things.

Skill

What Claude does

A work method. How to analyze, write, structure. The "how".

Example: a landing page audit Skill defines criteria, output format, tone.

MCP Server

What Claude accesses

A connection to an external service. The "where" and "with what".

Example: the PostHog server gives access to analytics data.

The real power is the combination. A reporting Skill + a PostHog server + a Slack server = a complete workflow. The Skill defines the method (which numbers, what format, which channel). The servers provide access (read PostHog, post to Slack).

Without Skills, MCP servers are useful but disorganized. You can ask PostHog for numbers, but you'll re-explain the format every time.

Without MCP servers, Skills are limited. You can have the best analysis method, but if Claude can't access the data, you're still copy-pasting.

Limits to know

Each server has its own permissions. A read-only server (PostHog, Notion) is safe. A server with write access (Slack, GitHub) can act on your behalf. Always check what you're authorizing.

Claude asks for confirmation before acting. By default, Claude shows what it's about to do and waits for your go-ahead before posting a message, creating an issue, or modifying a file. You keep control.

MCP servers don't communicate with each other. Each server talks to Claude, not to other servers. Claude connects the dots. This means orchestration depends on the quality of your instructions (CLAUDE.md + Skills).

The ecosystem evolves fast. New servers appear every week. The MCP protocol is an open standard, adopted by Google, OpenAI, and others. What works today will still work in 6 months.

Where to start

1. Identify your 3 daily copy-pastes

Which tools do you shuttle between every day? PostHog and Slack? Gmail and Notion? That's where MCP servers have the most impact.

2. Install 2 servers

No more. One to read (PostHog, Notion, Gmail), one to act (Slack, Google Sheets, GitHub). You'll see the difference immediately.

3. Combine with a Skill

You have a PostHog server and a Slack server? Create a "weekly report" Skill that reads analytics, formats the numbers, and posts them. A complete workflow in one command.

What's next

MCP servers connect Claude to your tools. The next article goes further: agents and sub-agents, which allow Claude to delegate complex tasks in parallel.

L’AI.ssentiel, every Friday

The AI signals that matter, a tool, a repo. For pros who already use AI and want to go further.

The AI signals that matter, the workflows that work, the shortcuts nobody explains. Every week, in your inbox.

Related articles

Where to start

The next level: from prompt to full workflow

Skills, MCP, agents, deployment. A structured path in 3 phases.