MCP's 97 million installs show agent infrastructure is settling around boring standards

Anthropic’s Model Context Protocol has reportedly passed 97 million monthly SDK installs only 16 months after launch. Treat the exact number with some care, because package installs are an imperfect proxy for production usage. Even so, the direction is clear enough: the agent market is starting to standardise around a shared way for models to talk to tools.

That is more important than another model leaderboard. Agents do not become useful because a chat window can produce better prose. They become useful when they can inspect a repository, query a database, open a ticket, read a calendar, update a document and do those things through interfaces that developers can reason about. Before MCP, every team was nudged towards a familiar mess: bespoke connectors for each model, each tool and each runtime. That does not scale. It produces fragile glue code and a security review for every new combination.

MCP’s appeal is that it is boring in the right places. It uses a protocol layer to separate clients from servers, so a tool provider can expose capabilities once and multiple AI clients can consume them. The Being Guru report points to more than 10,000 public MCP servers, broad client support across Claude, ChatGPT, Gemini, Cursor and other tools, and the protocol’s move into Linux Foundation governance through the Agentic AI Foundation. If that ecosystem holds, MCP starts to look less like an Anthropic feature and more like integration infrastructure.

The USB-C analogy is overused, but not useless. Standards win when they reduce decision fatigue. Developers do not want to choose a different cable for every device, and they do not want to write a fresh connector every time a product manager asks whether the agent can also use Jira, Slack or Postgres. A common protocol lets teams spend more time on permissions, workflows and reliability, which are the parts that actually decide whether agents survive contact with production.

There is a risk, though. MCP adoption can create a false sense of safety. A standard transport does not solve authorisation, data minimisation, audit trails or prompt-injection risk. If an MCP server exposes too much power, the model still has too much power. If a tool response includes hostile instructions, the client still needs to treat it as untrusted data. If every internal system becomes reachable through one agent surface, the blast radius may expand rather than shrink.

That means the practical work now moves up a layer. Teams should catalogue which MCP servers they allow, pin versions, review scopes, log tool calls and separate read-only capabilities from write operations. They should also build small, purpose-specific servers rather than dumping an entire internal platform behind one convenient endpoint. The standard makes integration easier. It does not remove engineering judgement.

For founders, MCP lowers the cost of making a product agent-compatible. For infrastructure teams, it offers a route away from one-off integration sprawl. For incumbents, it creates pressure to expose clean tool surfaces rather than hoping every AI client will implement their proprietary API directly.

The agent stack is still young, but this is how durable layers begin: not with spectacle, but with a boring protocol that everyone quietly decides is cheaper than the alternative.

---

Published: 2026-05-04 · Source: Being Guru