What Is MCP? A Team Guide for 2026

Model Context Protocol (MCP) is an open standard that Anthropic created in November 2024. It lets AI assistants like Claude connect to outside tools and data through one shared interface. Teams use it to give agents live context. SoWork's MCP server is one example: through it, Claude can ask who's around, what happened in a meeting, or what the team decided.

What Is Model Context Protocol (MCP), and Who Created It?

MCP is an open standard that defines how AI applications connect to external tools, data, and services. Anthropic introduced it in November 2024, and the Linux Foundation's Agentic AI Foundation now governs it.

Anthropic engineers David Soria Parra and Justin Spahr-Summers created the protocol [1]. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation [2]. That matters if you're buying tools, because no single vendor controls the spec. An integration you build today should keep working across AI platforms. Think of MCP as a universal adapter: build one connector and any MCP-compatible assistant can use it.

What Problem Does MCP Solve?

MCP solves the "N times M" integration problem. Before it existed, every AI model needed its own custom connector for every tool, so the number of integrations kept multiplying and they broke often.

Say you wanted Claude in your CRM, ChatGPT in your ticketing system, and an internal agent in your docs. That used to be three separate builds. With MCP, each AI app implements the client side once and each tool implements the server side once. The second problem is context. A model knows its training data and whatever you paste in. It has no idea what happened in yesterday's standup unless something connects it to that information, and MCP is that connection.

How Does MCP Work?

MCP uses a client-server design. A host app like Claude Desktop runs MCP clients, and each client connects to an MCP server that exposes tools, resources, and prompts.

Messages use JSON-RPC 2.0. Local servers run over stdio, and remote servers use Streamable HTTP [3].

How Is MCP Different From APIs, Plugins, and Function Calling?

MCP doesn't replace APIs or function calling. It sits on top of them. Function calling is how a model asks for a tool, and MCP standardizes where tools come from and how they run [5].

In most cases an MCP server wraps an API you already have and exposes a well-described subset of it as tools [4].

Which AI Tools and Platforms Support MCP?

Claude, ChatGPT, Cursor, Gemini, Microsoft Copilot, and VS Code all support MCP. By December 2025 the ecosystem had passed 97 million monthly SDK downloads and 10,000 active servers [2].

OpenAI announced MCP support in its Agents SDK in March 2025, and Google added MCP to its Gemini agent stack later that year [6]. ChatGPT added MCP connections in September 2025 [7]. Business tools are also shipping their own MCP servers. SoWork's server, for example, works with Claude, ChatGPT, Codex, and Gemini CLI [8]. Most teams can connect their existing assistant without switching AI providers.

What Are the Main Benefits of MCP for Teams?

The biggest benefit is reuse. You connect a system once and every MCP-compatible assistant can use it, which cuts integration work and keeps you from getting locked into one AI vendor.

• Less custom code: one server instead of one connector per model.

• Vendor flexibility: you can switch or mix models without rebuilding integrations.

• Better answers: agents work from live data, not stale copy-paste.

• Shared context: every agent on the team reads from the same source of truth.

The timing matters. Gartner predicts 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5% in 2025 [10]. Agents that can't reach your data won't do much for you.

How Can MCP Give AI Assistants Context About Meetings and Decisions?

Connect your assistant to the place where meetings and decisions already live. With SoWork's MCP server, Claude can ask what happened in a meeting or what the team decided [9], and it answers from real transcripts instead of guessing.

SoWork's AI Meeting Memory records, transcribes, and summarizes meetings automatically. MCP makes that history something you can query. Ask Claude "what did we decide about the pricing page on Tuesday?" and it pulls the answer from the actual meeting record. Through the same connection it can also summarize yesterday's meeting and post updates to your channels [8]. Nobody has to re-paste notes, and agents act on what the team agreed to rather than on assumptions.

How Can MCP Let AI Agents See Who's Around and What's Happening?

A workspace MCP server lets agents check real-time presence before they act. SoWork's MCP/API lets Claude ask who's around [9], using the same presence data your team sees.

In SoWork, Presence Awareness shows who's free, focused, or in a meeting. Automatic Calendar Sync updates someone's status when a calendar event starts. An agent can use that to send a question to whoever is actually available instead of interrupting someone in deep work. SoWork also puts AI agents in the office as characters next to your human teammates, so you work with both in one place. Agent desks, which will show what each agent is working on and what it did, are coming soon.

What Are Common MCP Use Cases by Team?

Engineering teams use MCP to query code and issues from the editor. Product teams use it to turn decisions into specs. Ops teams use it to automate status reporting.

The best use cases share one trait: the answer exists somewhere in your systems, but finding it means digging through several tools.

How Do You Set Up an MCP Server and Connect It to Claude?

If the server already exists, you add it as a connector in your AI client, sign in, and approve which tools it can use. Building your own means writing a server with an official SDK, such as Python or TypeScript.

1. Pick a server, starting with the vendor's official one.

2. Choose the transport. Local servers run on your machine through a config file, and remote servers connect over HTTP.

3. On Claude Team or Enterprise, go to Organization settings > Connectors, click Add, and choose Custom > Web [11].

4. Have each person sign in with their own account instead of sharing one key.

5. Test with read-only questions first.

Claude connects to remote servers from Anthropic's cloud, so your server has to be reachable on the public internet [11].

Should Teams Use Existing MCP Servers or Build Custom Ones?

Use existing servers for any mainstream tool, and only build custom ones for internal systems nobody else covers. Vendor-maintained servers get security fixes and spec updates that you'd otherwise have to handle yourself.

Most small teams don't need to build anything at first. Official servers from the tools you already pay for usually cover 80% of what you need.

What Security and Privacy Risks Come With MCP?

The main risks are prompt injection, tool poisoning, credentials with too much access, and malicious servers. The model acts on whatever servers send it, so bad data can turn into bad actions.

These risks have been measured. The MCPTox benchmark tested 20 LLM agents against poisoned tools and recorded attack success rates as high as 72.8% [12]. A separate server study found that 7.2% of servers had general security vulnerabilities and 5.5% had tool-poisoning vectors [12]. Microsoft also warns about "rug pulls," where a hosted tool's definition changes after you approved it [13]. On privacy, anything a server returns ends up in the model's context, so keep each server's access narrow.

How Should Teams Govern and Monitor MCP Agents?

Give every MCP connection the same treatment you'd give a new hire's access: least privilege, a named owner, and logs someone actually reads. The base protocol doesn't cover enterprise auth, audit, or governance, so you have to add those yourself [14].

1. Keep a list of approved servers and pin their versions.

2. Use per-user sign-in, not shared API keys.

3. Require a human to approve write actions.

4. Log every tool call and review the logs weekly.

5. Give each agent a named owner.

Larger fleets usually put an MCP gateway in front to centralize auth and monitoring. Gartner expects 75% of API gateway vendors to have MCP features by 2026 [15].

What Are MCP's Current Limitations?

MCP is young and still changing. The spec gets revised regularly, security tooling lags behind adoption, and server quality varies a lot across public registries.

The spec has already shifted in practical ways. A 2025 revision deprecated SSE in favor of Streamable HTTP [6], so older servers may need updates. Connecting too many servers fills the model's context with tool descriptions, which can lead it to pick the wrong tool. Network rules can also get in the way: Claude's remote connectors can't reach servers behind a VPN or firewall [11]. Plan for some maintenance. This isn't set-and-forget infrastructure yet.

What Does It Cost to Adopt MCP?

The protocol itself is free and open source. What you actually pay for is the tools you connect, hosting for any custom servers, and the engineering time to maintain and secure them.

The cheapest route is a tool that ships its own MCP server. SoWork is free for up to 10 users, Basic is $6/user/month, and Premium is $15/user/month [16], and the MCP/API is live.

How Should Teams Evaluate MCP-Enabled Tools?

Look at five things: who maintains the server, how it handles sign-in, whether reading and writing are separated, which compliance certifications it has, and whether it holds context your agents actually need.

As a reference point, SoWork is SOC 2 certified, offers SSO on Premium and Enterprise, and has SCIM on Enterprise only.

What Real-World Examples Show MCP Delivering Value?

MCP launched with Block and Apollo as partners, along with the developer tools Zed, Replit, Codeium, and Sourcegraph [14]. Today it pays off most when agents need live company context.

ChatGPT uses MCP connections for deep research across private data and company knowledge [7]. Developers in Cursor and VS Code query repos and issue trackers without leaving the editor. Distributed teams connect SoWork's MCP server so their assistant can see who's in the office, summarize yesterday's meeting, and post updates to channels [8]. The pattern is the same each time. MCP delivers the most where useful context would otherwise stay in people's heads or in meeting recordings nobody rewatches.

How Should a Team Get Started With MCP?

Start with one painful workflow, one AI client, and read-only access. Prove the value there before you add write actions or more servers.

1. Pick one workflow that wastes time, such as re-explaining meeting decisions.

2. Find official MCP servers for the systems involved.

3. Connect them to one client, like Claude, with read-only scopes.

4. Assign an owner and turn on logging.

5. Run it for two weeks and measure the time saved.

6. Add write actions and more servers once you trust the setup.

SoWork's rollout guide suggests picking two workflows, connecting Claude through the MCP/API, and naming an owner for each [17].

Bottom Line

MCP is the open standard that connects AI assistants to the tools and data your team already uses. It's now backed by every major AI platform and governed by a neutral foundation.

Start small. Connect one assistant to the system that holds the context your team keeps re-explaining, keep permissions tight, and expand from there. For distributed teams, that system is often your workspace, which is why SoWork's live MCP/API lets Claude answer who's around, what happened in a meeting, and what the team decided.

Key Takeaways

• MCP is an open standard that Anthropic introduced in November 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025.

• Claude, ChatGPT, Gemini, Microsoft Copilot, Cursor, and VS Code all support MCP, so one integration works across assistants.

• MCP builds on APIs and function calling. It doesn't replace them. It standardizes how AI apps discover and run tools.

• Security is the real cost: prompt injection and tool poisoning are documented risks, so use least privilege, per-user sign-in, and human approval for writes.

• Use official vendor servers first, and only build custom servers for internal systems.

• Workspace context is high value: SoWork's MCP/API lets Claude ask who's around, what happened in a meeting, or what the team decided.

• Get started with one workflow, one client, and read-only access, then expand once you've measured results.

Related reading

• SoWork vs Remo Pricing in 2026: Which Virtual Office Is Actually Cheaper for a Daily-Use Remote Team?

• Distributed Team Culture: How to Build It Without an Office

• Virtual Office Startup Review: Is SoWork Worth It in 2026?

Sources

1. MCP was introduced by Anthropic in November 2024 and created by engineers David Soria Parra and Justin Spahr-Summers

2. Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation; 97M+ monthly SDK downloads, 10,000 active servers, client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, VS Code

3. MCP uses JSON-RPC 2.0 with stdio for local servers and Streamable HTTP for remote servers; servers expose tools, resources, and prompts

4. An MCP server is an adapter exposing a well-described subset of an API as tools; MCP is a layer on top of APIs

5. MCP sits one layer above function calling; function calling is how a model asks for a tool

6. OpenAI announced MCP support in its Agents SDK in March 2025; Google adopted MCP in Gemini later in 2025; SSE deprecated in 2025 spec revision

7. ChatGPT added MCP integration in September 2025, used for deep research and company knowledge

8. SoWork's MCP server can see who's in the office, summarize yesterday's meeting, post updates to channels, and works with Claude, ChatGPT, Codex and Gemini CLI

9. SoWork's MCP/API lets Claude or your own tools ask who's around, what happened in a meeting, or what the team decided

10. Gartner: 40% of enterprise applications will include task-specific AI agents by end of 2026, up from less than 5% in 2025

11. Claude custom connector setup path (Organization settings > Connectors); Claude connects to remote MCP servers from Anthropic's cloud, so servers must be publicly reachable

12. MCPTox evaluated 20 LLM agents with attack success rates up to 72.8%; 7.2% of servers had general vulnerabilities and 5.5% had tool-poisoning vectors

13. Hosted MCP tool definitions can be changed after approval ("rug pull")

14. Base MCP protocol does not solve enterprise authentication, audit, or governance; launch partners Block and Apollo plus Zed, Replit, Codeium, Sourcegraph

15. Gartner: by 2026, 75% of API gateway vendors will have MCP features

16. SoWork pricing: free for up to 10 users; Basic $6/user/month; Premium $15/user/month

17. SoWork rollout guidance: pick two workflows, connect Claude through SoWork's MCP/API, assign owners

18. OpenAI tried ChatGPT plugins in 2023 but the ecosystem never took off