MCP crosses 97 million installs: AI finally speaks the same language as your tools
On March 25, 2026, the Model Context Protocol — an open standard created by Anthropic — crossed the threshold of 97 million monthly installs. It is the fastest adoption curve in the history of AI infrastructure standards. By comparison, Kubernetes took nearly four years to reach a comparable deployment density. But the number alone doesn't tell the whole story: in March 2026, OpenAI, Google, Microsoft and Mistral all integrated MCP support into their AI agent tools. More than 5,800 MCP servers are available, covering CRMs, ERPs, databases, collaboration tools and business APIs.
What it changes for you
✦ The opportunity
MCP solves the number one problem of AI in business: connecting to existing tools. Until now, plugging an AI agent into your ERP, your CRM or your accounting required weeks of development per connector. With MCP, this integration becomes standardized, fast and interchangeable.
No more custom integrations
Before MCP, every connection between AI and a business tool required specific code. Block — the fintech behind Square — eliminated 340 custom connectors by deploying MCP. For an SMB using 5 to 10 tools daily, this is the difference between 3 months of development and 3 days of configuration.
Switch models without rewiring everything
MCP is AI-model-agnostic. Today you use Claude, tomorrow you switch to GPT-5.4 or Gemini 3.1? Your integrations stay in place. This is the end of the vendor lock-in that held back so many SMBs from adopting AI. You choose the best model for each use case with no migration cost.
AI that acts within your tools
An AI agent connected via MCP doesn't just answer questions. It can read your emails, create tickets in your management tool, update your CRM, generate reports from your database — all via a secure protocol with granular permissions. This is the shift from “chatbot†AI to “coworker†AI.
âš The risk
Security and access scope
An AI agent connected to your business tools via MCP can read and modify data. The risk is not in the protocol itself — which includes three layers of permissions — but in the configuration. A poor definition of access rights can expose sensitive data. Every MCP server must be configured with the principle of least privilege: the AI accesses only what it strictly needs.
Variable maturity of the ecosystem
The 5,800 available MCP servers do not all have the same level of quality. The official connectors — Google Workspace, Slack, GitHub, Salesforce — are robust and maintained. Community connectors for niche tools can be less reliable. Before plugging an AI agent into a critical tool, check the maturity of the MCP server and test it in a controlled environment.
Our recommendation
MCP is no longer experimental — it's an industry standard. Here is how to take advantage of it right now:
Identify your 3 most time-consuming tools
Which tools consume the most time on repetitive tasks? CRM, emails, accounting, project management? Check whether they have an official MCP server. Google Workspace, Slack, Notion, Salesforce, HubSpot, PostgreSQL — the main B2B tools are already covered. If your tool is on the list, the connection can be done in a few hours.
Launch a pilot on a concrete use case
Don't connect everything at once. Choose a precise flow — for example: “the AI reads my emails for the day, identifies urgent client requests and automatically creates tickets in my management tool.†A targeted pilot lets you measure the real gain in 2 to 3 weeks, with no risk to the rest of your system.
Get support for internal tools
For your specific business tools or your ERP, creating a custom MCP server requires technical expertise. Plan for 2 to 5 days per connector. The investment pays off quickly: an MCP connector is reusable with any AI model, unlike a custom integration that ties you to a single vendor.
In summary
Frequently asked questions
Going further
MCP architecture in detail
MCP uses JSON-RPC 2.0 as its communication format — a standard proven over 15 years. The architecture rests on three components: the hosts — your AI applications such as Claude or an enterprise copilot —, the clients — which handle protocol communication —, and the servers — which expose the capabilities of your tools. Upon connection, the server sends a manifest describing the available tools, the accessible resources and the possible actions. The agent dynamically discovers what it can do, with no hardcoded code. Transport works locally via stdio or remotely via HTTP with Server-Sent Events. Three layers of permissions — server, client, host — ensure granular access control.