The Model Context Protocol (MCP) is now the common way AI assistants connect to business software: Claude, ChatGPT, Gemini, Microsoft Copilot and the main coding tools all support it, and companies like Stripe and HubSpot publish their own official MCP servers. In 2026 the protocol also got a neutral home, a major stateless rewrite and single sign-on through identity providers. For a business, the question has moved from "will this standard last?" to "which of our systems should we expose, and to whom?"
Here’s where adoption stands as of November 6, 2026, with each fact traced to the announcement that made it.
A quick refresher: what MCP does
MCP is an open protocol that lets an AI application (the client) discover and call tools, read resources and use prompts published by an MCP server. A server can wrap your CRM, a database, a ticketing system or an internal API. The point is that you build the connection once and any compatible assistant can use it, instead of building a separate plugin for each AI vendor. Our introduction to MCP covers the concepts in detail, and our MCP server development team builds these servers for companies that want their own systems reachable from AI tools.
Illustrative example Fictional business, people and data. Not a real client, conversation or result.
How did MCP get here?
Anthropic introduced MCP in November 2024. The milestones since then that matter for business use:
| Date | Milestone | Source |
|---|---|---|
| May 29, 2025 | MCP integration becomes generally available in Microsoft Copilot Studio | Microsoft Copilot blog |
| Sep 8, 2025 | The official MCP Registry is introduced | MCP blog |
| Dec 9, 2025 | Anthropic donates MCP to the Agentic AI Foundation (Linux Foundation) | Anthropic, Linux Foundation |
| Dec 10, 2025 | Google launches managed remote MCP servers for its cloud services, in public preview | Google, via TechCrunch |
| Jan 26, 2026 | MCP Apps becomes the first official extension: tools can return interactive UI | MCP blog |
| Jun 18, 2026 | Enterprise-Managed Authorization extension is stable, with Okta first | MCP blog |
| Jul 28, 2026 | The 2026-07-28 specification ships, with a stateless core | MCP blog |
| Aug 22, 2026 | Maintainers publish a new roadmap | MCP blog |
When Anthropic announced the donation in December 2025, it reported more than 10,000 active public MCP servers and over 97 million monthly SDK downloads across Python and TypeScript. By the July 2026 spec release, the maintainers said downloads across the Tier 1 SDKs were "close to half-a-billion" a month. Those are the projects’ own figures, not independent audits, but they show the direction clearly.
Which AI platforms support MCP?
Every major assistant now acts as an MCP client. What differs is how admins control it.
MCP support in the main AI platforms
| Platform | How MCP shows up | Admin control |
|---|---|---|
| Claude (Anthropic) | Connectors directory, custom remote servers, Claude Code; MCP Apps on web and desktop | Owners add connectors to Team and Enterprise workspaces; Enterprise-Managed Authorization supported |
| ChatGPT and Codex (OpenAI) | Remote MCP tools reach ChatGPT through plugins; Codex and the desktop app connect to MCP servers directly | Workspace admins choose which plugins and tools are available; creating plugins with MCP is off by default on Enterprise |
| Microsoft Copilot Studio | MCP servers added to agents as tools (GA since May 2025) | Requires generative orchestration; Agent 365 adds agent governance, with Defender mapping MCP servers per agent announced for June 2026 |
| Google Cloud and Gemini | Google-managed MCP servers for services like BigQuery, Maps, GKE and Cloud Run | IAM deny policies on MCP tool calls, Model Armor screening and Cloud Audit Logs |
| Developer tools | VS Code, Cursor and others connect to local and remote servers | Per-project configuration; VS Code supports Enterprise-Managed Authorization |
How MCP shows up
- Claude (Anthropic)
- Connectors directory, custom remote servers, Claude Code; MCP Apps on web and desktop
- ChatGPT and Codex (OpenAI)
- Remote MCP tools reach ChatGPT through plugins; Codex and the desktop app connect to MCP servers directly
- Microsoft Copilot Studio
- MCP servers added to agents as tools (GA since May 2025)
- Google Cloud and Gemini
- Google-managed MCP servers for services like BigQuery, Maps, GKE and Cloud Run
- Developer tools
- VS Code, Cursor and others connect to local and remote servers
Admin control
- Claude (Anthropic)
- Owners add connectors to Team and Enterprise workspaces; Enterprise-Managed Authorization supported
- ChatGPT and Codex (OpenAI)
- Workspace admins choose which plugins and tools are available; creating plugins with MCP is off by default on Enterprise
- Microsoft Copilot Studio
- Requires generative orchestration; Agent 365 adds agent governance, with Defender mapping MCP servers per agent announced for June 2026
- Google Cloud and Gemini
- IAM deny policies on MCP tool calls, Model Armor screening and Cloud Audit Logs
- Developer tools
- Per-project configuration; VS Code supports Enterprise-Managed Authorization
A note on OpenAI: it has used several names for this (connectors, apps and now plugins), and its current documentation describes a plugin as a package around an MCP connection. The Apps SDK it launched in October 2025 is also built on MCP. The protocol underneath has stayed the same, which is the useful part.
Google’s managed servers launched in public preview with general availability expected in 2026. Check the status of the specific service you need in Google Cloud’s documentation before relying on it in production.
Which business apps publish official MCP servers?
The bigger shift in 2026 is on the server side: software vendors now ship their own MCP servers, so you don’t have to build one for every SaaS tool you use. Three examples, from their documentation:
- Stripe runs a remote server at
mcp.stripe.comwith tools to read and write through the Stripe API and search its docs. Some actions, such as refunds, require a human to confirm them in the Stripe Dashboard. Stripe also says that after November 30, 2026 its MCP server will stop accepting full-access secret keys and restricted keys without the "Agent" tag, so autonomous setups need agent API keys or OAuth. - HubSpot offers a remote MCP server with read and write access to CRM records such as contacts, companies, deals and tickets, read-only access to items like campaigns and website pages, and it excludes sensitive data properties.
- Shopify removed the catalog and cart tools from its Storefront MCP endpoint and moved those functions to its Universal Commerce Protocol endpoint, while the Customer Accounts MCP server stayed unchanged. A reminder that vendor servers change too.
The Enterprise-Managed Authorization launch also listed Asana, Atlassian, Canva, Figma, Granola, Linear and Supabase as servers supporting single sign-on through Okta, with Slack adding support.
What changed in the 2026-07-28 specification?
The July 2026 release is the biggest change to the protocol so far. In plain terms:
- Stateless by default. The connection handshake and the
Mcp-Session-Idheader are gone. Each request carries its own protocol version, client identity and capabilities, so any server instance can answer any request. Servers that need state pass an explicit handle (a cart ID, a draft ID) between tool calls. - Easier routing. HTTP requests now carry headers naming the method and tool, so gateways and firewalls can route and filter MCP traffic without opening the body.
- Stricter authorization. Clients must validate the token issuer, and Dynamic Client Registration is deprecated in favor of Client ID Metadata Documents.
- Deprecations with a runway. Roots, Sampling, Logging and the old HTTP+SSE transport are deprecated, but the maintainers committed to keeping them working for at least twelve months.
- A formal extensions framework. MCP Apps, Tasks and Enterprise-Managed Authorization are now extensions on top of the core.
The TypeScript, Python, Go and C# SDKs supported the new spec as Tier 1 SDKs at release, and the Ruby SDK reached Tier 1 on October 30, 2026.
What does broad MCP adoption mean for your systems?
The technical barrier is mostly gone. What’s left is deciding what to expose and under which rules.
One connection, many assistants. A well-built MCP server for your order system works with Claude, ChatGPT, Copilot Studio and a custom agent. That protects you from vendor churn, which in 2026 has been real.
Login belongs to your identity provider. With Enterprise-Managed Authorization, access follows group membership and roles in your IdP, and there’s one audit trail. If you use Okta, ask your MCP vendors whether they support it.
Read-only first. Google’s own example policy denies every tool not marked read-only. That’s a sensible default for any company: start with tools that look things up, then add write actions one at a time, with approvals for anything that moves money or changes customer records.
Prompt injection is still the main risk. Stripe’s docs tell users to turn on tool approval and be careful when mixing its server with others. An assistant that reads an email and can also issue refunds needs guardrails between those two capabilities. Our guide on integrating AI into existing software covers where those controls go.
MCP readiness for your company
Where we fit
We build MCP servers and the AI integrations around them for US companies, from our Austin, Texas headquarters, with an engineering team that has shipped software since 2012. Typical projects wrap an internal database, an ERP or a line-of-business app that no vendor will ever publish a server for, with permissions and logs designed in from day one. For automations that run without anyone chatting, the same tools can feed workflow automation.
If you want your systems reachable from Claude, ChatGPT or Copilot without handing over the keys to everything, tell us what you need to connect.