MCP, the Model Context Protocol, is an open standard that lets AI agents plug into apps and data in a consistent way. A connector is a ready-made link to one service, such as Gmail, Slack or QuickBooks, and many connectors are built on MCP. Together they are how agents like Claude Duet, OpenAI dots and Gemini Spark reach your accounts without clicking through every website by hand.
What is MCP?
Think of MCP as USB for AI. Before USB, every device needed its own cable and port. Before MCP, every AI product had to build a custom integration with every app. With MCP, an app builds one "MCP server" describing what it can do (search email, create a calendar event, read a file), and any compatible agent can use it.
Anthropic created MCP and in December 2025 donated it to the Linux Foundation's new Agentic AI Foundation (AAIF), with OpenAI and Block as co-founders (IT Brief). That turned it from one company's project into a shared industry standard.
- Active public MCP servers (Anthropic, Dec 2025)
- 10,000+
- Monthly MCP SDK downloads (Anthropic, Dec 2025)
- 97M+
- Donated to the Linux Foundation's AAIF
- Dec 2025
- Major assistants adopting it: ChatGPT, Claude, Gemini, Copilot
- 4
When it announced the donation on December 9, 2025, Anthropic reported "more than 10,000 active public MCP servers" and "97M+ monthly SDK downloads across Python and TypeScript".
What is a connector?
A connector is a specific, pre-built integration that grants an agent access to one service. When Claude Duet says it has 6,000+ connectors, or dots says 4,000+ apps, they mean these ready-made links. Turning one on usually means signing in to the service and approving what the agent may access.
- Your agent
- Connector (often an MCP server)
- The app's API
- Your data and actions
Connectors vs computer use
Agents have two ways to use an app:
| Connector | Computer use (browser) | |
|---|---|---|
| How it works | Talks to the app's API in structured data | Looks at the screen, clicks and types |
| Speed | Fast | Slow |
| Reliability | High | Breaks when sites change or show CAPTCHAs |
| Coverage | Only apps with a connector | Any website |
| Example | Duet's QuickBooks connector | Grok Bot working in a CRM with no API |
Most agents use connectors where they exist and fall back to the browser everywhere else. Our how personal AI agents work guide covers the full picture.
Which agents use MCP and connectors?
As of November 2026, based on official docs and launch coverage:
- Claude Duet: 6,000+ connectors and plugins via the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, Docusign, Canva and Shopify. Anthropic's help center says Microsoft 365 is not yet available. See our Claude Duet review.
- OpenAI dots: 4,000+ apps through OpenAI's plugin ecosystem, with permissions shared across ChatGPT and Codex.
- Gemini Spark: Gmail, Docs and Workspace built in, plus MCP support for third-party services.
- Grok Bot: connectors and MCP where available, a Composio plugin covering 1,000+ apps, and computer use for the rest.
- Meta Muse: connector categories for email, calendars, payments, health, smart home, dining and shopping, using the browser where no API exists.
- OpenClaw: "skills" defined in
SKILL.mdfiles, many from the ClawHub marketplace, plus MCP.
Smart home example
On September 17, 2026, Google Home opened early-access MCP control to third-party agents, naming OpenClaw, Claude and Hermes. Google drew a clear line: those agents cannot unlock doors, disarm alarms or open garage doors (TNW). It is a good example of a connector exposing some actions and deliberately withholding others.
What are skills?
A skill is a packaged set of instructions and scripts that teaches an agent how to do a task, rather than just giving it access. OpenClaw popularised SKILL.md files, and Grok Bot turns workflows you demonstrate into skills. Connectors give access; skills give know-how. Many agents use both.
What are the risks?
- Prompt injection through tools. Researchers have shown that content returned by one MCP tool can carry hidden instructions that steer the agent's use of another. Simon Willison has written about this class of problem (Simon Willison), and in April 2025 Invariant Labs showed a WhatsApp MCP setup where one injected message, or a "sleeper" malicious MCP tool, made the agent leak chat history and contacts (Invariant Labs). See prompt injection explained.
- Malicious skills and servers. In February 2026's ClawHavoc campaign, Koi Security's first audit found 341 of 2,857 ClawHub skills were malicious, and Antiy CERT later counted 1,184 that had ever appeared there. ClawHub added VirusTotal scanning afterwards (The Hacker News).
- The lethal trifecta. Private data, untrusted content and the ability to act externally are the risky combination. Connectors often supply all three.
- Lingering data. Disconnecting an app from dots or Duet does not erase what the agent already learned from it.
Why should non-developers care about MCP?
You will rarely see the letters MCP in a consumer app, but the standard shapes what your agent can do. Because ChatGPT, Claude, Gemini and Copilot all support it, a company that builds one MCP server can reach users of every major assistant. That is a big reason connector counts jumped into the thousands in 2026. It also means connectors are becoming portable: switching from one agent to another is less painful when both speak the same protocol. The flip side is that a single flawed or malicious MCP server can affect users across many agents, which is why reviewing what you connect matters.
How to use connectors safely
- Start smallconnect one service you need, not everything on offer
- Prefer official connectorsavoid unknown third-party skills and servers
- Check scopeschoose read-only where it is offered
- Keep approvals onrequire confirmation for sending, paying and deleting
- Review regularlydisconnect services you no longer use
The UK's ICO lists least privilege, scoped permissions and human approval thresholds as key controls for agents. Our AI agent safety and permissions guide turns those into settings. For developers curious about adding text messaging to an agent, smsaiagents.com covers SMS via MCP. Unfamiliar terms are defined in the personal AI agent glossary.



