What Are Model Context Protocol (MCP)?
An open protocol that lets AI applications, such as Claude, ChatGPT, Cursor, or other AI assistants, connect to external tools and data sources through a standard interface called an MCP server.
2-minute setup • No credit card required
What it means
The Model Context Protocol (MCP) was introduced by Anthropic in November 2024 as an open standard and has since been adopted by many AI applications and developer tools. Before MCP, every AI app needed a custom integration for every data source. MCP replaces that with a common protocol, often compared to a USB-C port for AI: build one MCP server for your system and any MCP-compatible client can use it. The architecture has three roles. The host is the AI application the user interacts with (a chat app, an IDE, a coding agent). Inside the host, an MCP client manages a connection to each server. The MCP server exposes capabilities, mainly tools (functions the model can call, such as 'search issues' or 'summarize activity'), resources (data the app can read), and prompts (reusable templates). Messages use JSON-RPC 2.0. Servers can run locally or remotely. A local server runs on the user's machine as a process the host starts, communicating over standard input and output; it is common for file system access or local developer tools and inherits the user's local permissions. A remote server runs as a web service reached over HTTP (the current spec's transport is called Streamable HTTP), is usually authenticated with OAuth or an API key, and can be shared by a whole team without installing anything. For Git data, MCP servers range from general-purpose ones, such as GitHub's official MCP server, which can read and often write repositories, issues, and pull requests, to narrower read-only servers that expose summaries and reports rather than raw repository access.
Why Model Context Protocol (MCP) matter
MCP makes it practical to ask an AI assistant questions about real, current company data instead of copying and pasting context. For engineering teams that means asking 'what shipped in the payments repo this week?' or 'which PRs are waiting on review?' directly in Claude, Cursor, or another client, and getting an answer grounded in live data. Because it is a standard, a team can connect the same server to several AI tools and switch tools without rebuilding integrations. The flip side is security: an MCP server gives a model the ability to take actions and read data with whatever permissions the server holds, so the choice of servers and their scopes matters as much as the choice of AI tool.
How to measure
MCP is a protocol rather than a metric, but adopting it safely can be evaluated with a short checklist. Scope: does the server need write access, or can it be read-only? Prefer read-only servers for reporting and analytics. Data exposure: does the server return raw source code or secrets, or only metadata and summaries? Authentication: is access tied to individual users or keys you can revoke, with per-key scopes and expiry? Trust: is the server from a vendor or maintainer you trust, and do you review tool descriptions, since a malicious or compromised server can try to steer the model (prompt injection)? Auditability: can you see which tools were called and when? Cost: some servers consume paid credits per call; check before connecting them to an agent that may call tools in a loop.
Real-world example
An engineering manager wants to prepare for a Monday planning meeting from inside Claude. She connects two MCP servers. The first is a local file-system server limited to a notes folder. The second is a remote, read-only Git reporting server authenticated with an API key that only grants access to her team's repositories. She asks: 'Summarise what shipped last week across our repos, list PRs opened by AI agents that are still waiting for review, and draft three discussion points.' Claude calls the server's activity and agent tools, gets structured results, and writes the summary. Because the Git server is read-only and never returns source code, the worst a misbehaving prompt can do is read summaries she already had access to. Gitmore's remote MCP server works this way, with read-only tools such as activity_summary, agent_ledger, find_changes, recent_activity, agent_reports, and list_repositories, and its tool calls do not use AI credits.
Common questions
What is an MCP server?
An MCP server is a program or web service that exposes tools, resources, or prompts to AI applications using the Model Context Protocol. The AI app (the host) connects to it through an MCP client and can then call its tools, for example to search a database or summarise Git activity.
What is the difference between a local and a remote MCP server?
A local server runs on your machine, started by the AI app and communicating over standard input/output, with your local permissions. A remote server runs as a hosted web service over HTTP, is authenticated with OAuth or an API key, and can be shared across a team without local installs.
What is a Git MCP server?
An MCP server that gives AI assistants access to Git data: repositories, commits, pull requests, issues, or summaries of activity. Some, like GitHub's official server, can also write (create issues, comment, open PRs). Others are read-only and return reports and metadata rather than code.
Is MCP secure?
MCP itself is a protocol; security depends on the servers you connect and how you scope them. Use read-only servers where possible, grant keys the minimum repositories and permissions needed, prefer revocable per-user keys, only install servers from trusted sources, and be aware that tool output can contain text that tries to manipulate the model.
Which AI apps support MCP?
Many do, including Claude (desktop, web, and Claude Code), and a growing list of IDEs, coding agents, and AI assistants from other vendors. Support for remote servers and authentication varies, so check each client's documentation.
Track Model Context Protocol (MCP) Automatically
Gitmore turns your git activity into automated reports with real metrics — delivered to Slack and email.
Get Started FreeNo credit card • No sales call • Reports in 2 minutes