TL;DR: The Model Context Protocol went from an Anthropic side project to the standard adopted by OpenAI, Google, and Microsoft in under a year. It's the protocol that lets AI assistants actually do things — not just talk about doing things. Here's why it matters and what it looks like when you connect it to build infrastructure.
In November 2024, Anthropic open-sourced a protocol called MCP — the Model Context Protocol. The pitch: a standard way for AI assistants to connect to external tools and data sources. No more bespoke integrations. No more pasting API responses into chat windows. Just a universal interface between AI models and the systems they need to interact with.
Fourteen months later, MCP is the de facto standard for AI connectivity. The official servers repository has over 73,000 GitHub stars. Monthly SDK downloads crossed 97 million — a 970x increase from launch. Over 10,000 active MCP servers are in the wild.
How MCP Won
In March 2025, OpenAI adopted MCP across the Agents SDK, Responses API, and ChatGPT. Two weeks later, Google followed with Gemini support. By May, Microsoft made MCP generally available in Copilot Studio. Within six months, the three largest AI companies in the world had adopted a protocol created by one of their competitors.
The reason: network effects. Every MCP server that gets built — Stripe for payments, Sentry for error tracking, Atlassian for project management — adds value to every AI assistant that supports the protocol. The flywheel is spinning, and opting out of MCP means opting out of the ecosystem.
In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation — co-founded by Anthropic, OpenAI, and Block, with support from Google, Microsoft, and AWS. MCP is no longer one company's project. It's vendor-neutral infrastructure.
What MCP Actually Is
Think of it as USB for AI. Before USB, every peripheral had its own connector. MCP does the same thing for AI-to-tool connectivity: any service that implements an MCP server exposes structured capabilities — tools, data queries, guided workflows — to any AI assistant that speaks the protocol. Cursor, Claude, Windsurf, GitHub Copilot, and ChatGPT all support it natively.
Before MCP, connecting an AI assistant to your internal tools meant custom integrations, API keys in prompts, and hoping the model could parse unstructured output. MCP replaces that with a single standard. The assistant discovers available tools, understands their parameters, and calls them programmatically.
The Highest-Leverage Use Case: Developer Infrastructure
AI coding assistants already understand your code. They can read source files, parse dependency graphs, and generate patches. But until MCP, they were blind to what happened after you wrote the code. They couldn't see your builds, query your test results, or trace a failure through your remote execution cluster.
That's why we built Hermetiq's MCP server. It exposes your entire Bazel build intelligence layer — everything we capture from the Build Event Protocol — as structured tools AI assistants can call. Not a chatbot wrapper around logs. A full programmatic interface to your build telemetry: invocation traces, cache miss analysis, remote execution breakdowns, flaky test detection, infrastructure health metrics.
What Changes
A developer pings you in Slack: "My build is broken." You open the CI run, scroll through thousands of lines of Bazel output, cross-reference the dependency graph, check the remote cache — forty-five minutes later, you find a misconfigured visibility rule three levels deep in a transitive dependency.
With Hermetiq's MCP server, the developer opens Claude or Cursor and asks: "Why did my build fail?" The assistant queries the build telemetry, inspects the invocation trace, identifies the failing action, compares it against recent successful builds, and explains the root cause — in under a minute.
"Our cache hit rate dropped this week." → The assistant pulls 30-day trends, identifies the inflection point, finds a toolchain update that changed the compiler flags hash, and recommends the fix.
"Which tests are costing us the most time?" → The assistant ranks tests by execution time multiplied by frequency, identifies three tests accounting for 40% of total test time, and flags two as flaky.
"Is our remote execution cluster healthy?" → The assistant checks scheduler queues, worker utilization, and storage latencies in parallel. Finds two workers in a restart loop due to OOM kills. Recommends the memory limit fix.
Where This Goes
Today the integration is interactive — a developer asks, the assistant investigates. But the same tools an AI assistant calls interactively can be called by an autonomous agent. The same telemetry that helps a developer understand a failure can trigger an agent that fixes it automatically.
MCP won because it solved the right problem at the right time: giving AI assistants hands, not just a mouth. The teams that connect their infrastructure to it will ship faster and debug less. The teams that don't will keep scrolling through logs. Connect Hermetiq's MCP server and see the difference.
See what MCP + build intelligence looks like. Explore Hermetiq's product to see the dashboards, cache analytics, and cost tracking that power our MCP server.
Related Articles
- We Built Hermetiq to Solve the Problem Created by the AI Coding Boom — The build infrastructure bottleneck and Hermetiq's vision for self-healing builds.
- What Is Build Observability? — Metrics, logs, and traces for understanding what happens inside your Bazel builds.
- What Is Hermeticity and Why It Matters — The foundational property that makes caching and remote execution reliable.