The limitation of AI without integrations is that it can only work with what you bring to it. You copy text from HubSpot, paste it into Claude, get a response, copy the response, paste it somewhere else. Every step is a manual handoff. That friction doesn't disappear at scale — it multiplies.
Model Context Protocol (MCP) changes this. It lets Claude connect directly to the tools and databases in your existing tech stack — so instead of copying data in and out of conversations, Claude can read from your CRM, write to your database, run searches in your SEO platform, and post to Slack, all within a single workflow. The AI goes from a tool you use to a layer that coordinates your systems.
At Prometheus, we use MCP daily. Claude connects to Ahrefs for keyword research, to Supabase for our content database, and to Slack for automated notifications from our content performance engine. This article explains what MCP is, how we use it in practice, and how to set it up for your own tech stack.
What MCP Is, in Plain English
MCP (Model Context Protocol) is an open standard, published by Anthropic in November 2024, that defines how AI models communicate with external tools and data sources. Think of it as the connection layer between Claude and everything else in your environment.
Before MCP, every integration between an AI model and an external tool required custom code: a one-off API connector, a specific plugin, a manual pipeline. MCP standardizes that connection so that a compatible "MCP server" — a small piece of software that wraps your tool's API — can be plugged into Claude without custom integration work for each combination.
Anthropic's protocol documentation described the purpose directly: "MCP is a universal, open standard for connecting AI assistants to the systems where data lives — including content repositories, business tools, and development environments."
The analogy that sticks: MCP is to AI tools what USB is to peripherals. Before USB, every device needed its own connector. After USB, one standard let everything plug in. MCP is doing the same thing for AI-tool integrations.
Shawn Wang (known as swyx), a prominent AI engineering writer and co-founder of Latent Space, called MCP "the most consequential infrastructure release in the AI tooling ecosystem in 2024" in his end-of-year analysis. "Not because of what it does today, but because of the surface area it opens for the next two years of AI agent tooling."
Our Actual MCP Setup at Prometheus
Here's how MCP shows up in our real daily workflows — not as a demo, but as the actual mechanism powering things we do every day.
Ahrefs MCP: SEO Research in Context
We have an Ahrefs MCP server configured in our Claude environment. When we're doing keyword research — like the research behind this very article — Claude can call the Ahrefs API directly. Keyword difficulty scores, search volume, related terms, competitor traffic estimates: all available within a Claude session without leaving the environment to manually pull data from the Ahrefs dashboard.
The keyword research for this content cluster ran entirely through Claude + Ahrefs MCP. We pulled difficulty scores and volume for terms like "claude code mcp" (4,000 monthly searches, difficulty 11), "cursor vs copilot" (2,400 monthly searches, difficulty 9), and "claude cowork" (44,000 monthly searches, difficulty 46) in a single session. What previously required switching between multiple tabs and copying data into a doc happened in one conversation.
Supabase MCP: Database Access for Content Operations
Our content platform runs on Supabase. Through the Supabase MCP server, Claude can read and write to our CMS database directly — checking existing content coverage before writing a new article, verifying that internal links point to real slugs, querying which posts need quality fixes, and checking glossary term coverage. This is the integration that enables our autonomous content pipeline to function: Claude can interrogate the database without a human running queries.
IDC's 2025 AI Integration Survey found that 68% of enterprise AI projects cite data access friction as their primary barrier to productivity. MCP is the practical answer to that friction — it's what makes AI systems capable of operating on real business data rather than data you manually copy into a chat window.
Slack Integration: Automated Notifications from AI Workflows
Our content performance engine — which monitors keyword movement, content quality scores, and AI crawler activity — posts automated reports to Slack via MCP. When our glossary pipeline publishes a new batch of terms, a Slack notification fires automatically with a summary of what was created and what links were injected. When a post drops in ranking, the performance engine flags it.
This is the workflow pattern that turns AI from a tool you use into infrastructure that keeps you informed. The human doesn't need to pull reports — the system surfaces what changed.
How to Set Up MCP Servers
Setting up MCP requires editing a configuration file that Claude reads at startup. For Claude Code (the terminal agent), the configuration lives at ~/.claude/config.json. For the Claude desktop app, it's in the application settings under "Developer." The format looks like this:
{
"mcpServers": {
"ahrefs": {
"command": "npx",
"args": ["-y", "@ahrefs/mcp-server"],
"env": {
"AHREFS_API_KEY": "your_key_here"
}
},
"supabase": {
"command": "npx",
"args": ["-y", "@supabase/mcp-server-supabase"],
"env": {
"SUPABASE_URL": "your_url",
"SUPABASE_SERVICE_ROLE_KEY": "your_key"
}
}
}
}Each entry in mcpServers names a tool and defines how to run its server process. Most MCP servers are available as npm packages, which means setup is typically a one-line install and an API key. Claude automatically discovers the available tools when it starts and can call them by name in any session.
Most Useful MCP Servers for Business Teams
The MCP ecosystem has grown fast since Anthropic published the standard. As of early 2026, there are official MCP servers for the following business-relevant tools:
Data and analytics: Google Analytics, BigQuery, PostgreSQL, MySQL, Supabase, Snowflake. If you have a data warehouse, Claude can query it directly and build analyses without a data engineering middleman.
CRM and marketing: HubSpot, Salesforce (via community server), Ahrefs, Google Search Console. The HubSpot MCP integration is particularly useful — Claude can read contact records, create deals, update properties, and trigger workflows without leaving the AI session. For teams doing AI + CRM integration work, this is a significant capability.
Communication and project management: Slack, GitHub, Linear, Jira, Notion. Connecting Claude to your project management stack means the AI can create tasks, update statuses, and post summaries from within workflows rather than requiring human orchestration between systems.
Developer infrastructure: Filesystem access, Docker, AWS, Cloudflare. For engineering teams, these enable Claude Code to manage infrastructure as part of its autonomous workflows — not just reading and writing application files but interacting with the deployment environment.
HubSpot's marketplace alone lists over 500 native integrations, and McKinsey's 2025 enterprise technology survey found that the average mid-market company uses 56 distinct SaaS applications. The opportunity that MCP represents is clear: instead of each tool requiring its own AI integration, one protocol lets Claude work across all of them.
Building Connected AI Workflows
The real value of MCP isn't any single integration — it's what becomes possible when Claude can touch multiple systems in a single workflow. Here's an example from our own operations:
When we identify a new content opportunity, Claude can: (1) run keyword research via Ahrefs MCP to validate search volume and competition, (2) query our Supabase database to check whether we already have coverage for that topic, (3) check our glossary terms table to identify related terms to cross-link, (4) generate the content brief, and (5) post a Slack notification to the content team with the opportunity summary and brief attached. That's a five-step workflow across four systems, orchestrated by a single Claude session.
Without MCP, each step requires a human to manually retrieve data from one system, format it, and pass it to the next step. With MCP, Claude handles the coordination. The human's job becomes reviewing the brief and making the call to proceed — not the data plumbing in between.
If you're building AI agent capabilities more systematically, our guide to creating AI agents for business covers the broader architecture. For the CRM-specific integration patterns, the AI + CRM integration playbook goes deep on HubSpot and Salesforce specifically.
Getting Started With MCP
The fastest path to a working MCP setup is to identify the one tool your team touches most frequently and start there. For most businesses, that's either the CRM (HubSpot or Salesforce), the database (any SQL-compatible system), or a marketing tool (Ahrefs, Google Analytics).
Requirements: a Claude Pro or Max subscription (for the desktop app) or an Anthropic API key (for Claude Code), and API credentials for the tool you're connecting. Installation is typically three steps: find the MCP server npm package, add it to your config file, and restart Claude.
Stanford HAI's 2025 AI in the Enterprise report found that organizations with AI tools connected to existing systems see 2.4x higher productivity gains than organizations using AI tools in isolation. The pattern tracks with what we've seen internally: isolated AI assistance produces incremental gains, connected AI workflows produce step-change gains.
For a broader framework on where to begin with AI adoption, see our AI readiness assessment guide. If you want support configuring MCP for your specific tech stack, our AI Enablement team handles this as part of the AI integration engagements we run for mid-market clients.
