Skip to content
tl;dr of MCPs
MCP is an open standard that lets AI models read data from tools like Google Drive, Slack, and meeting recorders like tl;dv. Anthropic released it in November 2024 and donated it to the Agentic AI Foundation, under the Linux Foundation, in December 2025.
MCP standardizes how access is requested and granted, but it does not govern what a server does with your data once it arrives.
The current specification, 2026-07-28, made the protocol stateless and hardened its OAuth authorization. AWS, Google Cloud, Microsoft, Cloudflare, Figma, Netlify, Supabase, and Xero have all shipped support.
tl;dv runs a read-only MCP server on Pro plans, listed in the Claude and ChatGPT connector directories.
What is MCP in AI?
Model Context Protocol (MCP) is a way for AI models to link in with your apps and tools. Think of it as a universal adaptor for AI. Just like how USBs are the standard method of connecting hardware, MCP is now a standardized method of connecting software to LLMs.
It used to be that many models were stuck in their own little worlds. They couldn’t easily interact with real-time information from your apps, like Google Drive, Calendar, or Zoom. That’s where Model Context Protocol (MCP) comes in. Developed by Anthropic, MCP is how AI connects with external tools, making it more actionable and easier to access rather than daisy-chaining across multiple tools and apps.
One way to look at it is to imagine that every time AI connects to a new tool, it has to work in a new language. This makes the Model Context Protocol the new standard for ‘translating’ different tools. In other words, AI doesn’t need custom connectors for each and every tool anymore. It simply uses MCP as a layer to access the information it needs from the necessary tool.
Whether it’s summarizing your emails, pulling relevant documents from Google Drive, or automating meeting insights, MCP enables AI to work with your existing workflows.
With MCP, AI models can now access real-time information from external sources.
More and more enterprise AI solutions are adopting MCP because it simplifies integrating AI with everyday tools. For example, an AI could pull data from your CRM system, analyze it, and then use that data to send you personalized recommendations or alerts.
Why Did Anthropic Create MCP?
Model Context Protocol was developed by Anthropic, the AI research company behind Claude, in November 2024. The goal? To make AI smarter, more useful, and safer when working with real-world data. Since then, the governance has moved to the Linux Foundation as of December 2025.
Why is MCP Popular in Enterprises?
For enterprises, MCPs make AI models much more practical for real-world applications. Companies want AI that can plug into existing tools like their CRM systems, cloud storage, or calendar apps without custom workarounds each time. MCP allows that connection to happen.
Businesses have already adopted MCP for tasks like automating reports, analyzing customer data, and organizing team workflows. Instead of building complicated and costly integrations, they can now use MCP to bring together data from different sources and marry it up with other context you have from within the business.
Model Context Protocol, more commonly known as MCP, is the open standard that lets AI models read data from the tools you already use.
Anthropic released it in November 2024. Nearly two years on, it’s running in production at AWS, Google Cloud and Microsoft, and the current specification shipped in July 2026. This guide was updated in August 2026 to reflect it.
Below we break down what MCP is, how it works, what it’s good at, and where the protocol stops and the vendor’s responsibility starts.
We also cover tl;dv’s own MCP server, which is live and listed in both the Claude and ChatGPT connector directories. With your permission, your AI app can search your meeting library, pull transcripts and read the notes, without you opening a single recording.
If you’d like to hear an expert explain MCP instead, be our guest:
What are the Main Benefits of the Model Context Protocol?
MCP has five benefits: standardized communication, security and authentication, real-time processing, multi-platform capability, and finally, customization.
1. Standardized Communication
First is how MCP standardizes AI models’ communication with external tools and platforms.
- Without MCP, AI models would need custom coding for every single connection.
- With MCP, it creates a universal language that works across different APIs, databases, and platforms.
This means that an AI model can easily connect to and retrieve data from tools you commonly use, whether it’s Google Drive, Zoom, Salesforce, or something else entirely.
2. Security and Authentication
MCP specifies how an AI client requests access and how a server checks it, using an authorization framework built on OAuth. That means permissions are scoped to what you approve and are revocable.
3. Real-Time Processing
MCP can perform real-time processing. This means that when you need data from a tool (say, Google Calendar for scheduling), the AI model doesn’t have to wait around for outdated information.
4. Multi-Platform Compatibility
MCP’s functionality is its ability to work across multiple platforms. Whether you’re using Google Workspace, Slack, Zoom, or even something like Microsoft Teams, Model Context Protocol means that you can call up information from all of these separate entities into a single place.
5. Customization
MCP is built to be customized and adapted for different use cases. This customization is ideal if you’re building an automation workflow for your marketing team or creating a personalized AI assistant for your business.
This makes it perfect for both small-scale applications and larger, enterprise-level solutions that require AI to handle complex workflows across many different systems. In short, MCP is a fully customizable layer that allows you to connect AI with tools you want in real-time. To take that and turn it into action is where agentic workflows come in.
How Does MCP Work?
MCP works by having a common format for two pieces of software to communicate with each other. First is a client inside your AI app, and second is a server in front of the tool that holds the data.
The Request/Response Flow: What Is It?
The request/response flow is the core of the Model Context Protocol. Here’s how it works in practice:
- The model asks for something
- The client sends the request to the server
- The server sends a response
- The client returns the data to the model
1. The Model Makes a Request
For example, you’re working on your calendar and need the AI to check your schedule. The AI sends out a request, kind of like a little note saying, “Hey, can you grab my calendar events for today?”
Without MCP, that request means nothing to your specific tool. It needs to be turned into a language that the tool recognizes.
2. The Client Sends the Request to the Server
The AI app’s MCP client puts together the request in a standard format and sends it to the tool’s MCP server. Because the two sides both understand the same format, this allows the server to read the request without the need for custom code to be written for that particular app.
3. The Server Sends a Response
The calendar tool (or whatever app you’re using) checks its data, finds the relevant events (say, a meeting at 2 PM), and sends the info back to the client. This is the response stage, where your calendar tool delivers the requested data.
4. The Client Returns the Response to the Model
Once the client receives the response, it passes the information back to the AI app, which hands it to the model. At this point, the model has the data it needs (your schedule for the day) and can process it however it’s been programmed to. In this case, maybe it gives you a little reminder or even suggests prepping for a 2 PM meeting.
This works behind the scenes, making sure the AI and external tools are speaking the same language.
Real-World Example: AI Summarizing Your Meeting Insights
You’ve just finished a Zoom meeting that was packed with ideas and action items. You’re busy, your brain’s fried, and now you have to sift through your meeting notes. Enter AI agents for meetings and MCP.
- AI Requests the Zoom Meeting Data
The AI requests the meeting recording and transcript. The request goes to the meeting tool’s MCP server, which handles the Zoom side. - The Model Gets the Data and Analyzes It
Once the AI has the meeting data, it can start analyzing it. Maybe there’s a list of action items, a few questions, and some insights shared by your team. The AI summarizes all of that, highlights the key points, and organizes them for you in an easily digestible format. - AI Responds with a Summary and Next Steps
Instead of you having to go through pages of meeting notes, the AI delivers a quick, summarized report along with prompts for the next steps. It might even ask, “Would you like me to schedule follow-up meetings based on these action items?” and boom, your meeting is turned into an actionable, organized list.
In this example, MCP was the shared format that let the AI app and the meeting tool understand each other. The retrieval came through that connection, and the summarizing was the model’s own work.
How Does MCP Handle Security?
MCP standardizes how an AI app seeks access to a tool and how that access is granted. What happens to your data after the server receives it is decided by whoever runs the server and not by the protocol.
How MCP Handles Authentication
MCP’s authorization framework builds on OAuth. The AI app requests a defined set of permissions; you then approve these, and the server then issues a token ringfenced to what you granted. Your password is never handed to the AI app, and you are able to withdraw the access later without changing any passwords.
The 2026-07-28 specification, the largest revision since MCP launched, hardened this further. Authorization servers now return an issuer parameter that clients must validate before redeeming a code, and client credentials are bound to the issuer that minted them, so they cannot be reused across authorization servers.
Where MCP Ends and the Vendor Begins
MCP governs the request itself. The organization running the server governs the data.
Once a server receives a request, the protocol has no say in how long the data is kept, whether it is used to train a model, or who inside the vendor can see it. Those are decisions made by whoever operates the server, in the same way they already are for every other tool your business uses.
Example: Connecting an AI App to Google Drive
Say you need your AI app to source some documents stored in your Google Drive before a big client meeting. The app requests a specific set of permissions for the Drive, and you approve them using your Google account. Each request the app makes after that permission is granted has a token limited to what is approved, and Google’s own systems choose what that token can reach.
MCP’s part ends there. What the app does with the documents once it has them, whether it keeps a copy, how long it holds it, comes down to whoever built the app.
Choosing a Server
Prefer first-party servers run by vendors you already have a contract with. Check who operates the server and what their published data policy says. Servers from public registries need more care, since the operator is often harder to trace.
For larger organizations, the Enterprise-Managed Authorization extension became stable in June 2026. It lets administrators provision MCP access centrally through an identity provider rather than leaving each person to approve servers one at a time, and Anthropic, Microsoft, and Okta have adopted it.
What Can MCP Do for a Sales Team?
MCP allows an AI model to pull live CRM records, email threads, and meeting transcripts at the moment you ask, rather than working from whatever you upload or paste into the prompt.
The result is that you can ask the AI about a specific deal or client in your pipeline, rather than working with a generalized picture.
This is already running at real volume. In the July 2026 specification release, Honeycomb reported that close to 20% of their monthly interactive queries now come from agents rather than people.
Where It Shows Up in a Sales Workflow
A rep is about to join a client call. Rather than opening four tabs from the CRM, call library, and other sources, the model can fetch everything in one place. The email exchange, the deal’s current status in the CRM, and notes from the previous call.
The same access can also support the admin that follows on from the call. Meeting notes from the AI notetaker are logged to HubSpot or Salesforce. A follow-up email is drafted, grounded in what was actually discussed on the call. Deal stages can be updated. A stalled deal can be flagged with a quick summary of what happened before it went quiet.
According to Salesforce, sales teams can lose up to 72% of their time due to admin, so the practical use of AI in this situation is recouping that lost time. This is the ground that AI agents for sales are built on.
Who’s Actually Using MCP?
MCP is past the experimental stage. AWS, Google Cloud, Microsoft, Cloudflare, Figma, Netlify, Supabase, and Xero all shipped support for the current specification in July 2026.
Developers building MCP connections use ready-made toolkits, one for each programming language. Those toolkits are downloaded close to half a billion times a month, and two have passed a billion downloads in total.
What That Looks Like in Practice
There are two really good examples of this, outside of the engineering aspect.
Customer Support
An AI support agent, connected over MCP, can do things like check an order’s status in Shopify or pull customer tickets from Zendesk. The reply is then built from the account’s actual history rather than a pre-determined template.
Project Management
AI agents for project management work across several tools at once. The model reads the week’s meetings from Google Calendar, checks the related documents in Notion, and posts a summary to Slack, all from one request.
How Can MCP Be Applied to Meetings?
The ability to query and pull in real meeting transcripts is one of the most useful things that can be connected to a model. They are verbatim records of conversations, decisions, objections, and actions that can sometimes slip through the gaps of manually typed CRM notes or documents.
Which of the AI meeting assistants you use matters here, since not all of them run a server.
What AI Can Do With Meeting Data
The AI model, when querying the transcript library, can pull in the full record of what was said. With that it can:
- Summarize a call you didn’t attend, without needing to watch the recording.
- Pull action items across multiple calls, so you can get the bigger picture of the status.
- Find when a decision was made, and what the context was around it.
Here’s how that works on three of the major platforms.
MCP for Zoom
Client calls and external meetings mostly happen on Zoom. Connect a library of those calls to a model, and you can ask what a client objected to three calls back without opening any of them. Our Zoom MCP guide covers the setup.
MCP for Google Meet
Meet sits inside Workspace, so the documents around a meeting live next to the meeting itself. A model reaching both can tie what was agreed on a call to the doc that came out of it. More detail in our Google Meet MCP guide.
MCP for MS Teams
Teams is usually where the internal meetings happen, which makes it the biggest library of the three for most companies. Querying across it surfaces recurring topics and decisions rather than one meeting at a time. Our Microsoft Teams MCP guide goes further.
tl;dv’s MCP Server
tl;dv has run its own MCP server since 2025 and, like MCP itself, it has changed a lot in that time. It is now available on the Pro plan and above, and connects from both the Claude and ChatGPT connector directories. The connection uses OAuth, so you approve a scoped set of permissions and can revoke them at any time without changing your password. Tokens last five minutes.| At launch | As of August 2026 |
|---|---|
| API key, generated in settings and pasted into a config file | OAuth, with scoped permissions you approve and can revoke |
| Installed locally with Node.js or Docker | Hosted by tl;dv, nothing to install |
| Config file edited by hand, then restart your client | Connect from the Claude or ChatGPT connector directory |
| Four tools | Five tools, including search across your meeting library |
- Search meetings: filter by date range, keywords, internal/external type, participation, and pagination.
- Get meeting metadata: retrieve title, date, duration, status, organizer, participants, and meeting URL.
- Get meeting transcript: retrieve the full transcript in Markdown.
- Get meeting notes: retrieve AI-generated notes, manual notes, timestamps, and AI-processing status.
- Get user profile: retrieve your tl;dv profile, role, registration date, and team size.
Where MCP Goes Next
In 2026, expectations for MCP have moved far beyond the previous status quo. Models rely on it heavily now, security has been tightened, and setup doesn’t require as much technical prowess. Before, you needed to edit config files. Now you select from the connector directories and it runs.
With the July 2026 revision, authorization rules became stricter, and Enterprise-Managed Authorization became stable in June 2026, so administrators can provision access through their identity provider rather than approving servers one at a time.
The direction from here is less dramatic. The 2026-07-28 revision made the protocol stateless, added a formal extensions framework, and introduced a twelve-month deprecation policy.
That framework brought two official extensions. MCP Apps lets a server return an interactive interface rather than plain text, rendered in a sandboxed window, with every action routed through the same approval path as a normal tool call. Tasks handle work that takes a while, so a server can hand back a job reference instead of holding the connection open.
Together, it means MCP can now grow without breaking what people have already built, which is what infrastructure looks like once it stops being interesting.
FAQs About Model Context Protocol (MCP)
What is Model Context Protocol (MCP)?
MCP is an open standard that lets AI models connect to the tools and data you already use, like Google Drive, Slack or your meeting recordings. Anthropic released it in November 2024 and governance has since moved to the Linux Foundation. It gives every AI app one common way to reach external data.
How is MCP different from AI agents?
They work together rather than competing. An AI agent is a system that carries out tasks on your behalf. MCP is how that agent reaches the data it needs to do the work. Most useful agents rely on something like MCP to see beyond their own training data.
Can an MCP server change or delete my data?
It depends on the server. MCP supports tools that retrieve data and tools that take actions, so what a server can do is decided by whoever built it. Check which tools a server exposes before connecting. Read-only servers can fetch information but cannot create, edit, share, or delete anything.
Is MCP safe to connect to my work tools?
MCP standardizes how access is requested and granted, using OAuth so permissions are scoped and can be revoked. What happens to your data after a server receives it is decided by the company running that server. The safety question is less about the protocol and more about which server you connect.
How do I know whether an MCP server is trustworthy?
Start with who operates it. First-party servers from vendors you already have a contract with come with a data policy and someone accountable. Check what permissions it requests and what its retention policy says. Servers from public registries need more care, since the operator can be harder to trace.
Does connecting MCP send my data to ChatGPT or Claude?
Data retrieved through MCP enters the AI app’s context so the model can use it, which means the privacy policy of whichever AI you’re using applies. The MCP server, the AI app, and the model provider are three separate parties, so it’s worth reading the terms for each.
Can I use MCP to search my meeting transcripts?
Yes, if your meeting tool runs an MCP server. tl;dv’s is available on Pro plans and connects from the Claude and ChatGPT connector directories. It exposes search, metadata, notes and transcripts, so you can ask questions across past calls rather than opening them one at a time.
