The Current State of Cloud MCP Servers
20 months on: MCP is the default way AI acts in your apps. The official registry passed 67,000 servers, the current spec adds async tasks and skills, and Zapier, Make, Composio and n8n all reshuffled. What changed, what it costs, and what to actually pick.
When we wrote our first take on the MCP landscape in April 2025, the Model Context Protocol was eighteen months old and still finding its feet. It was Anthropic's experiment that OpenAI had just said it would adopt; Zapier was the only vendor with a serious hosted offering; the "official registry" did not exist yet.
Twenty months on, the picture has changed completely. This is a refresh of that post: what the protocol actually is today, what the numbers really say, where the major players moved, and what we bet on at Mazaal AI.
The protocol, as it stands now
MCP shipped in November 2024 and has been revised four times since: 2025-03-26, 2025-06-18 (which moved the standard transport to streamable HTTP), 2025-11-25, and the current revision, 2026-07-28 (modelcontextprotocol.io, checked 15 and 19 November 2026). The current spec is a deliberately thin core — JSON-RPC 2.0, stateless requests, per-request capability negotiation — with optional extensions that both sides negotiate at startup:
- Tasks — asynchronous execution for long-running operations, with polling, mid-flight input and durable handles.
- Skills over MCP — rich, structured instructions for agent workflows, discovered through the protocol itself.
- MCP Apps — interactive UI (charts, forms, video) rendered inline in a conversation.
The security section also grew up. Our 2025 post complained that no spec version treated tool descriptions as untrusted; the current spec now says exactly that — annotations "should be considered untrusted, unless obtained from a trusted server" — and requires explicit user consent before any tool invocation.
The numbers (and their honest fine print)
The official MCP registry launched in September 2025 with 407 servers. The MCP team's own figures: close to 2,000 entries by 25 November 2025 (a 407% increase from launch), "over 10,000 published servers" by the project's donation to a foundation in December 2025, and 9,652 latest-version records as of May 2026 (devtoolhub.com, gyld.ai). A full pull of the registry API on 10 September 2026 returned 30,375 unique servers from 99,114 version records (devtoolhub.com) — roughly 3× the May count, with August 2026 alone adding more servers than the registry's entire first five months. 98.9% are marked active; almost nothing is ever deleted. The growth has not slowed in proportion: direct full pulls of the same API on 19 November 2026 (14:00 and 18:36 UTC) returned 67,595 and 67,703 unique server names — 66,608 marked active, 1,095 deprecated at the later pull — more than double the September count in 70 days. Between the two pulls, 108 new servers appeared in four and a half hours: net additions are running at roughly 500–600 per day — far below the September–October surge that doubled the registry, but the growth has not plateaued.
Third-party directories count differently: Smithery listed 11,245 and PulseMCP tracked 21,982 (1 September 2026, agentscamp.com); mcp.so self-reported around 23,000 (July 2026). GitHub returned 15,926 repositories tagged mcp-server in May 2026, and the canonical modelcontextprotocol/servers repository holds 86,148 stars.
The fine print matters more than the headline:
- 62% of registry servers have exactly one published version — shipped once, never revised. Roughly 36% haven't been updated in three months or more (10 September pull).
- Most of it is moving, though. In the 19 November pulls the median server's latest version is 36 days old; 45.0% published a version in the last 30 days and 75.1% in the last 90. The stale tail is real — 24.9% of servers show no update in 90+ days and 9.3% in 180+ — but smaller than the 36% in the September pull, diluted by the growth itself.
- Distribution is top-heavy. About 18,200 publishers; the top 50 account for 24.6% of all servers, and the largest single publisher owns 1,505 single-function micro-tools (10 September pull). The 19 November pulls say the same in sharper form: the three largest publishers alone hold 9.5% of all server names (2,366 / 2,099 / 1,936), and the top ten for 15.8%. The raw count is inflated by automated publishing operations, not tens of thousands of hand-built integrations.
- The registry verifies the namespace, not the server. Publishing proves you own
io.github.your-accountorcom.yourdomain. There is no security-review field, no liveness check, and 22.9% of servers link no source repository at all (37% of remote-only ones). "Official" means "run by the MCP project" — not "vetted". - Servers are not apps. One server can expose dozens or thousands of actions — Zapier's single endpoint serves 66,000+ triggers and actions. So "30,000 MCP servers" is not "30,000 integrations", and not "30,000 things your agent can usefully do".
The players, twenty months later
Zapier MCP — the 2025 critique was architectural: one URL, every tool in the context window. Zapier changed the design. The current product (zapier.com/mcp, 15 November 2026) gives the agent discovery tools instead of a loaded tool list — it searches 9,000+ apps and 66,000+ triggers/actions for what fits. When no standard action fits, Code Actions let the agent write and run its own code against an internal API. Work that proves out is saved as a reusable Skill or deployed as a "Next Gen Zap" (early access; running them is free until 1 December 2026). The scale claims on the page: 150,000+ people running agent-led actions, 500,000+ MCP servers created, 25,000,000 tool calls, 1.3 million apps connected. It's included in every Zapier plan at two tasks per tool call, with action-level permission controls and SOC 2 Type II. The breadth play has clearly won.
Make MCP — the April 2025 bet that existing Make scenarios would become MCP tools is now the product (make.com/mcp, 15 November 2026). Any number of scenarios can be exposed as callable tools over 3,500+ apps, with you deciding which parameters the AI controls and which stay constant; GDPR, SOC 2 and SOC 3 on the page. The 2026 addition that matters: an MCP Client too — Make scenarios and Make AI agents can now consume external MCP servers (monday.com's, among others), so Make sits on both sides of the protocol. AI can even manage the Make organisation itself (folders, teams, invites).
Composio — the April 2025 snapshot said "up to 250 tools". In 2026 their hosted MCP serves 1,500+ apps to Claude, Cursor, ChatGPT and Codex-class clients (bitdoze.com, 2026), with user-scoped connection URLs and API-key enforcement per connection. Their own published position is the sharpest one in the ecosystem: server count is a vanity metric; the real criteria are governance, versioning, and security enforcement.
n8n (new since the 2025 post) — the self-hosting favourite now plays both roles (n8n.io/mcp, docs.n8n.io, 2026): an MCP Client Tool node (landed late 2025) lets any AI Agent node consume external MCP servers, and an MCP Server Trigger plus an instance-level MCP server expose your workflows — with tools for creating, validating, testing workflows and generating test data. This is the "your automations become agent tools" path without adopting a hosted platform.
Cloudflare and Auth0 — our 2025 assessment (remote servers on Workers, OAuth via workers-oauth-provider; Auth0's identity layer for remote MCP) was not re-verified for this refresh; treat it as the 2025 state of those two. The self-hosted-remote-server path they pioneered is still how most "bring your own server" setups work.
The platform wave. The 25 November 2025 one-year post (blog.modelcontextprotocol.io) is the best single snapshot of who adopted what: OpenAI (MCP "integrated across ChatGPT and our developer platform"), Google (Gemini, Gemini CLI, open-source servers for Google Maps and Cloud databases), Microsoft (Foundry went "from a small set of tools to thousands" across GitHub, Azure and M365; GitHub CPO on the server, registry and enterprise allowlists), AWS (Bedrock, Kiro, Strands, AgentCore, Quick Suite), Block (goose, Square AI, Moneybot), Okta (Cross App Access as an authorization extension), Hugging Face (HF-MCP, Gradio). The MCP team's own list of servers to try: Notion, Stripe, GitHub, Hugging Face, Postman. One telling detail from the September 2026 registry pull: the founding vendors have published almost none of their servers to the registry they back — com.microsoft: 16, com.anthropic: 0. Discovery for the servers that matter still runs through docs and word of mouth.
Security, one year on
Our 2025 list of open risks — compromised server abusing delegated scope, malformed tool descriptions tricking agents, no risk classification in the spec — is half-closed. The 2026-07-28 spec now mandates treating tool annotations as untrusted and explicit consent per invocation, and extensions are opt-in by both sides. What the spec still cannot do is vet a server for you. The practical filters from the registry data:
- Start from the namespace — a verified-domain namespace is a company staking its domain; a single GitHub account with 1,000+ servers is a bulk operation. Judge the individual server.
- Check version and last-updated, not the description. "1.0.0, eight months old, no repo link" is a skip, especially for a remote server your agent talks to over HTTP.
- Follow the repository link before you connect anything. For a remote-only server with no repo, that is a hard stop for production.
- Re-verify auth against the server's own docs — the registry will not tell you whether it implements the current OAuth 2.1 flow.
Mazaal's take (updated)
Our original thesis was that the ecosystem felt like TCP/IP before routers — every network speaks the protocol, but they can't find each other — and we built against three gaps: discovery (a registry where any MCP server advertises its capabilities), a policy envelope (auth, rate limits and cost controls around every tool call), and a negotiation layer between incompatible auth flows.
Twenty months later, the first piece is shipped and the framing has sharpened. What is live on mazaal.ai today (checked 15 and 19 November 2026): a 1,597-app direct catalog, plus an MCP Extensions surface on the /apps page that reaches thousands more tools through MCP providers. Discovery is no longer a bet — it's a tab. The policy envelope and the negotiation layer remain the roadmap; every month of spec churn makes both more necessary, because the current spec deliberately leaves auth heterogeneity and per-tool accounting to the host.
The endgame we published in 2025 still reads right: agents that evolve, not configurations you re-wire. Yours starts using a new PDF parser because the directory shows it's faster, and drops it when something better ships. In tools, as in markets, liquidity beats ownership — and the registry, whatever its noise, is the exchange.
FAQ
What is MCP (the Model Context Protocol) in 2026?
An open, open-source protocol — announced by Anthropic in November 2024 and now backed by OpenAI, Google, Microsoft, AWS, GitHub and others — that standardizes how LLM applications connect to external tools, data and actions over JSON-RPC. The current revision is 2026-07-28 (modelcontextprotocol.io, 15 November 2026).
How many MCP servers exist in 2026?
The official registry held 30,375 unique servers as of 10 September 2026 (99,114 version records; devtoolhub.com) and 67,703 unique server names — 66,608 active — as of a direct full pull at 18:36 UTC on 19 November 2026 (67,595 a few hours earlier in the day). Third-party directories report 11,000–23,000 depending on who counts and how. The honest read: supply is not the bottleneck — 62% of servers have exactly one version, and while 46% were updated in the last 30 days, a quarter show no update in 90+ days.
Is the official MCP registry vetted or safe?
It verifies that a publisher owns the namespace they publish under, but it stores no security review, no auth details and no liveness check, and 22.9% of entries link no source code. Treat a listing as "this exists", not "this is safe" (devtoolhub.com, 10 September 2026).
Are MCP servers the same as apps or integrations?
Not 1:1. A server is a protocol endpoint; one server can expose dozens or thousands of actions (Zapier's single endpoint serves 66,000+ triggers and actions, zapier.com/mcp, 15 November 2026). So a "12,000+" or "30,000" server count is a measure of endpoints, not of integrations or of useful actions.
Sources and check dates
- modelcontextprotocol.io — specification (current revision 2026-07-28, re-checked 19 Nov) and changelogs; blog.modelcontextprotocol.io "One Year of MCP" (25 Nov 2025) — fetched 15 Nov 2026.
- registry.modelcontextprotocol.io — direct full API pulls, 19 Nov 2026 (version=latest ×3 at 14:00 / 14:50 / 18:36 UTC, updated_since, all-version records): 67,595 / 67,621 / 67,703 unique names, 66,608 active at the 18:36 pull, staleness and publisher statistics above. First-party; no second-party intermediary.
- devtoolhub.com "The MCP Registry, By the Numbers" — full API pull, 10 Sep 2026 — fetched 15 Nov 2026.
- agentscamp.com "MCP Ecosystem Statistics 2026"; gyld.ai "MCP Adoption Statistics 2026" (May 2026 figures) — fetched 15 Nov 2026.
- zapier.com/mcp; make.com/mcp — fetched 15 Nov 2026.
- bitdoze.com Composio guide (2026); n8n.io/mcp and docs.n8n.io MCP server tools reference — fetched 15 Nov 2026.
- mazaal.ai/apps and mazaal.ai/pricing — fetched 15 Nov 2026.
- Cloudflare MCP and Auth0 paragraphs: 2025 assessment, not re-verified for this refresh.
Keep reading
All articles →
AI
AI Agent Examples: How Businesses Leverage AI for Growth (2025)
AI agents are intelligent software programs capable of autonomously performing tasks, making decisions, and adapting to real-time data. Unlike traditional automation tools that operate on rigid rules, AI agents leverage machine learning and natural language processing to understand context and optimize outcomes.

AI agents
AI Agents: Your New Virtual Team Members for Smarter Business Automation
Imagine if you had a tireless assistant who could handle routine tasks, answer questions, and even make minor decisions for you—all without needing a coffee break. Sounds like science fiction, right? AI agents make this a reality. Read more on how AI agents are changing the world.

AI agents
What we shipped: May 2026
Agent Skills replace Prompts, the dashboard gets one consistent design with real metrics on workflow cards, and the agent builder and conversations dashboard now work properly on mobile.