AI agents · 6 min read

MCP at 18 Months: 97M Installs, 17K Servers

MCP went from launch to 97mn monthly SDK downloads in 16 months — 970× growth that beat React's three-year curve. Scraping vendor differentiation now compresses to one axis: appearing in the agent's tool list when it asks for capabilities.

By Signal Census Editorial
Apify
Apify · MCP at scale
Apify's MCP server exposes 20,000+ Actors as discoverable tools — the largest single tool surface available to any MCP-compatible LLM client in 2026.
Bright Data
Bright Data · MCP at scale
Bright Data's MCP server exposes proxies, Web Unblocker, datasets, and Deep Lookup as agent-callable tools — the largest enterprise-tier scraping surface in MCP.

On March 25, 2026, Anthropic’s Model Context Protocol crossed 97 million monthly SDK downloads. Sixteen months earlier, that number was roughly 100,000. The 970× growth curve is faster than any AI infrastructure standard has hit before — and faster, by a clean margin, than React took to reach 100 million monthly downloads, which required roughly three years.

The headline number is the easy part. The harder question is what the curve does to businesses that previously sold scraping infrastructure as REST APIs with vendor-specific SDKs. The answer is now visible in the data, and it is not subtle.

The registry filled in 12 months

A year ago, the public MCP server registry held about 1,200 servers. By April 2026, it had crossed 9,400+ public servers, with month-over-month growth still tracking at +18% across Q1 2026. An independent census from Nerq indexed 17,468 servers across all registries combined — roughly double the public count, once private and enterprise-internal servers are included.

For comparison, the earlier 5,000-server figure that anchored coverage as recently as March 2026 is now outdated by more than 3×. Growth is not slowing in the relevant timeframe.

The shape of what is in those 17,468 servers matters more than the count. Roughly three categories dominate: tool wrappers around existing SaaS products (Slack, GitHub, Linear, Notion), data-source connectors (databases, file systems, vector stores), and scraping/extraction surfaces from vendors who used to sell that same capability as a REST API. The third category is where the competitive math reshapes.

Enterprise has already moved

Concrete enterprise penetration numbers are messy — most public estimates come from vendor surveys with their own methodology issues. The directionally consistent finding across recent vendor-published surveys (Anthropic, Vucense, Digital Applied) is that MCP adoption among enterprise AI teams has crossed the majority threshold during 2025 and continues to rise. The “should we ship an MCP server?” conversation closed several quarters ago for vendors selling into AI-agent buyers. The remaining question is what tool surface to expose, and how to make it the one the agent picks.

Governance moved out of Anthropic

In December 2025, Anthropic donated MCP to the Linux Foundation, establishing it under the Agentic AI Foundation (AAIF). Platinum members at launch included AWS, Google, Microsoft, Cloudflare, and Bloomberg.

The governance shift matters for one reason: it neutralised the competitive concern that adopting MCP meant betting on Anthropic specifically. Once Google, AWS, and Microsoft sit on the foundation that owns the spec, no buyer needs to evaluate MCP as an “Anthropic protocol.” It is the substrate. Native MCP support now ships in Claude, ChatGPT (via Apps SDK and Connectors), Google Gemini API and Vertex AI Agent Builder, Cursor, Windsurf, Zed, JetBrains AI Assistant, the Vercel AI SDK, and the OpenAI Agents SDK. The list is not exhaustive; it is the floor.

For the scraping vendor selling into AI-agent buyers, that is the actual competitive frame. The buyer is no longer choosing between scraping APIs the way they chose between REST APIs in 2022. The buyer is choosing between tools that appear in their MCP-compatible client’s tool list, ranked by the agent’s tool-selection logic.

What this changes for the scraping stack

The economics implication is sharper than the integration story.

When an LLM agent invokes a tool through MCP, three things happen that did not happen under the old SDK model. The agent reads the tool description, reasons about cost-versus-benefit, and executes. All three steps are runtime decisions made by the model rather than design-time decisions made by a developer. Vendor lock-in built around “we already integrated X” evaporates because the integration is one connection string. The remaining moat is whatever the agent observes about success rate, latency, and price during its selection step.

Three concrete shifts follow.

Apify’s distribution math improved. The Apify MCP server exposes 20,000+ Actors as runtime-discoverable tools. The historical drop-off curve — browse Store, read README, sign up, configure UI, run — collapses to a single typed tool call. With x402 (USDC on Base) and Skyfire payments, the agent transacts without ever creating an Apify account. For any Actor whose value parses cleanly from its input schema, the addressable demand expands meaningfully.

Bright Data’s small-surface bet looks better than it did. The Bright Data MCP server exposes a tight set of high-success-rate tools rather than a discoverable catalog. AIMultiple’s benchmark ranked it first at 76.8% task success, 48.7 seconds average completion. That is the design choice that wins when an agent’s tool-selection prompt is cost-constrained: fewer tools listed means fewer tokens consumed during selection, and a documented success rate beats a longer catalog with unknown reliability per tool.

The publishers who lose are the ones with long READMEs. The Q1 2026 lead-extractors census showed that demand in that segment is already organised around a single positioning phrase — “No Cookies” — that fits in the actor title. MCP compresses that pressure another order of magnitude. An Actor whose pitch lives in a marketing landing page or a long README is invisible to an agent that reads only the structured tool metadata. Names, schemas, and one-paragraph descriptions become the entire surface.

The token-cost question is still open

The unresolved frame from the prior MCP coverage on this site has not been resolved. When an agent calls a scraping API through MCP, it first reads the tool list (potentially thousands of entries), then reasons about which to call, then formats and parses the result. All of that is LLM token spend, and it sits on the buyer side of the ledger.

For the foundation labs — Anthropic, OpenAI, Google — that token consumption is pure revenue. For the scraping vendors, it is a tax on every transaction that flows through their MCP surface. The per-call rate the vendor charges has to clear the marginal token cost of having the agent route to them rather than around them. That is a tighter equilibrium than the REST-API era required.

Two paths are visible. Vendors competing on the smallest, highest-quality tool surface — Bright Data’s approach — minimise the token cost of agent selection and keep their per-call rate competitive against the alternative. Vendors competing on catalog breadth — Apify’s approach — bet that agent-side pre-filtering (showing the agent only the top-N relevant tools per query) will keep the discoverable subset narrow at runtime even when the catalog is enormous.

Both bets are plausible. Which one captures the relevant share of the browser-agent demand stack is the question the next four quarters will answer.

What 970× growth in 16 months actually means

The React comparison is worth one more look. React reached 100 million monthly downloads after roughly three years of being the canonical way to build a frontend application. MCP reached 97 million monthly downloads after sixteen months of being one of several proposed agent-integration standards. The difference reflects, at minimum, two things: the size of the AI-agent total addressable market relative to the frontend developer market, and the absence of competing standards once Anthropic open-sourced the spec.

The practical read for any vendor selling into agent-built workflows is that the standardisation battle is over. There is no second-place protocol contender to wait out. The question is which tool surface wins inside the standard — and that is where the competition for scraping infrastructure dollars happens for the rest of 2026.

Sources