What is MCP? The full guide

[.blog-callout]

✨TL;DR:

• MCP lets AI assistants connect to external apps and data, so they can find information and take actions on your behalf.
• It standardizes these connections, making it easier to give AI access to tools across multiple apps without building each integration from scratch.
• Softr supports MCP for both builders and app users, letting AI work with Softr workspaces or published apps while keeping access controlled by permissions.

[.blog-callout]

You know that more context produces better AI responses. So you start the hunt: you gather documents, round up files, and cram everything into a crafted prompt. But as you spend more time searching and lining everything up, you're now working for your AI agent, not the other way around. You're losing the driver's seat.

MCP puts the controls back in your hands. You can use it to connect your favorite agents to software you use, so AI can search and securely manipulate your data on your behalf. No code or lengthy setup involved.

What is MCP?

MCP (Model Context Protocol) is an open standard that gives AI agents a common way to connect to external tools and data. It lets AI do two important things: pull in the context it needs and take permitted actions in the software you use. Created by Anthropic, it replaces custom connectors with a single interface, making the integration process simple and easy to maintain. This is why it's often referred to as the USB-C of software integrations: a single protocol that covers every app.

Why is MCP important?

AI assistants are powerful, but most business context lives somewhere else—in your CRM, inbox, project tools, databases and documents. And answering a question is only half the job. To actually get work done, AI needs a way to interact with those systems.

MCP provides a standard way to do both. Instead of creating a different integration between every AI assistant and every application, MCP gives compatible systems a common way to expose their data and capabilities.

For starters, it's much easier to set up for non-technical people. With MCP, you simply select the app you want to connect to, log in to your account in the external tool to approve permissions, and then it's ready to use.

Before MCP, you'd have to configure each individual capability in the chatbot interface or with code. Even if you were connecting to the same app such as Asana, you'd have to add a tool to read your projects and another one to create a new project. MCP just loads all the capabilities when the session starts, so you get access to the full list without the setup complexity.

Since app developers update their apps often, MCP saves you from having to actually go back into each tool to edit the details to comply with what's new. You don't have to do maintenance on the connector, so you can spend more time working on what's important, not on figuring out why your integration doesn't work every few months.

And since it's an open standard, more apps are now launching their MCP servers. The universe of possible connections keeps increasing, placing your AI agent at the center, letting you interact with all of them.

The protocol's popularity is taking off, going from a niche power tool to a mainstream feature. This is both thanks to how MCP has evolved and also tied to how newer AI models are becoming much better at what's called agentic tasks (the capacity to work with multiple tools and stay focused on a goal until it's complete to an acceptable degree).

How does MCP work?

MCP exposes the apps you use to your AI assistant, so you can access and act on them using natural language commands.

Here's an overview of how it works.

Finding MCP connectors

Apps like ChatGPT, Claude, or Cursor all have an MCP client, letting them send requests to external servers. The external apps, such as HubSpot or Stripe, provide MCP servers that process those requests and send the results back to your agent.

ChatGPT has a Plugins page with a full list of all supported MCP servers. When you open the details page, you can see the actions that the server supports. Not all apps allow full control over creating, updating, or deleting resources, so be sure to check to see if it has the capabilities you need.

As for Claude, clicking the Connectors tab shows you all the native MCP integrations available. The information page also details what capabilities are included. On top of that, you can upload a file to set up a connector not present on the original list, so you can connect to niche software you need to work with.

Connecting for the first time

Once you select an MCP server from the list, the first step is logging into your account on the external platform and approving permissions. For example, if you're connecting ChatGPT to Notion, you'll be taken to Notion's login page and then through a dialog showing what privileges ChatGPT will have.

After finishing the authentication process, both apps are now connected and you can start a new chat to explore the new connector. For security reasons, you might be asked to log in again every few days or months, but the day-to-day is frictionless.

The data transmission between both platforms is secure; you can't see data from any other user that uses this external app, and no one else can see yours at any time.

MCP capabilities

When you start a new chat and use an MCP connector, both the server and your AI app remain connected for the entire session. The server exposes the range of capabilities it offers, similar to offering a restaurant menu to your agent. These can include:

• Tools, letting you create, read, update, or delete data, as well as running a process or workflow in the external tool you're connecting to. You can use these by telling the agent to take action on the external app (for example, "Mark my project as complete in Todoist")
  
• Resources, which are read-only documents provided by the server that you can select and add as context to the agent. You can ask the agent directly to read these documents.
  
• Prompts, which are instruction templates for interacting with the external platform.
  

MCP servers might have any combination of the above; the developers of the external tool decide which capabilities each MCP server has, which is tied directly to the app's features. Read the integration page to understand what you can and can't do.

If you ask to do an action that the server doesn't support, the agent will often tell you that it can't execute it; in very rare cases, it can go into a loop and try to find workarounds to get it done. If you find that your agent has been working for a long time without showing results, check whether the connectors lack the functionality you need.

How to know when you're using an MCP server

Depending on the agent you're using and your settings, you have to activate an MCP connector before using it in a new chat thread. This is usually done by selecting it from a dropdown menu in the prompt input field, typing a slash command, or explicitly naming the app you want to connect to (for example, "get my latest Gamma presentation").

Accessing external data is a sensitive action, so most AI apps will ask before starting to use an MCP server. A permission request box will display in the chat thread or main interface, with options to allow or deny access.

If you set it to always allow until the end of the session, you can check the thought trace of the tool to see what it's accessing. This is the text that changes after you send your prompt, showing whether the model is thinking, searching the web, or using a tool.

Connecting multiple MCP servers to the same agent

You can connect multiple MCP servers to the same AI assistant, exposing it to a wide range of tools and resources. You can, for example, connect your Claude app to Google Drive, Figma, and Canva; the agent will always have a list of capabilities for all these apps at the same time.

As you chat, when the agent detects an intent or command that matches a connected MCP server, it sends out a request to the appropriate app. The external app receives and processes the request, sending back the response to the agent. You can now use the result of that response as context: for example, using the content of a Google Docs to create a new infographic on Canva.

There are caveats to connecting multiple MCP servers to the same agent. If you connect apps with similar features, the agent may get confused when using tools. For example, if you have both Google Drive and Dropbox connected, asking for your contract templates without telling the agent specifically which cloud drive it's on can have it try a random one.

If you keep facing situations where the agent is using the wrong tool, consider being more specific ("Find the contract templates in my Dropbox") or temporarily disabling an MCP server.

MCP vs API

If you've been building your own apps with visual development tools, you might have come across the API standard (application programming interface) for connecting to external platforms. By writing a few lines of code or setting them up in a visual interface, you can push or retrieve data into external apps from any app you're building.

MCP doesn't replace the API standard. In fact, it works as a translation layer on top of it: the MCP server routes the requests sent by agents to the respective API endpoints, and then returns the results once they're ready. If the API is updated at any time, the MCP server will offer new capabilities without any extra setup on your side.

What can you do with MCP?

The MCP standard is so flexible that users keep finding new ways of connecting agents and apps together. Combined with increasing model capabilities and more robust MCP servers, there's a lot to experiment with as you streamline and automate your work.

When most of your work happens in ChatGPT, you can integrate the MCP servers of the other apps you use more often. For example, when connecting to Google Workspace:

• You can ask for a breakdown of the top email threads in your inbox
• Talk about any file in your Google Drive, and have it adapted, repurposed, or transformed in any way you need
• Check your calendar by asking questions in natural language so you're on top of what's coming up on your schedule

Since you can use multiple MCP connectors in the same chat thread, you could even turn an email into an event on your calendar, or save an attachment from a message to your cloud drive.

Beyond chatting directly with your AI assistant, you can use scheduled tasks (available in both ChatGPT and Claude) to automate actions across apps you use. For example, you can connect the Notion MCP server and get a weekly summary of tasks complete, or a daily digest of Intercom customer service tickets handled. The agent fires at the day and time that you set, and executes the prompt using the connected tools and MCP servers to give you the response you want.

More complex use cases involve orchestration, where multiple systems come together to produce a response. Sales is a good example:

• You have a CRM with client data, enrichment services to look up information for new leads, and a wide range of company systems that hold product and service information, along with rules on how to build deals.
• Reps can use AI to enrich a contact, build a draft deal proposal, and get it ready to send to the client in minutes, all without leaving the same chat thread.
• The AI agent connects to each app, looks up the necessary information, and then keeps the CRM up to date, automating a lot of the manual work.

All of these steps require only natural language. You can still access the apps in the browser to check that the actions were performed correctly, but you'll find yourself doing it less once you start trusting your agent.

How Softr uses MCP

Softr supports running your own MCP server, so your AI assistants can connect directly to the apps you've built on Softr.

Softr supports MCP in two ways, depending on whether you want an AI assistant to help you build and manage your Softr applications or let people interact with an app you've built.

The Softr MCP Server connects an AI assistant such as ChatGPT, Claude, or Cursor to your Softr workspace. This lets you use natural language to work with your databases, apps, integrations, and workflows. For example, you could ask your assistant to search records, add a field to a database, make changes to an app, or create a workflow.

You control what the AI tool can access when you connect it. Permissions can be set separately for applications, databases, and workflows, with different levels of access depending on how much you want the assistant to be able to see or change.

Softr also has an App MCP Server, which works differently. Instead of connecting an AI assistant to the workspace for building and managing, it lets the people who use a published Softr app connect that app to their own AI assistant.

In this case, the assistant acts on behalf of the signed-in user and is limited by what that person can already access through the app. The app's existing user groups, page visibility, and data restrictions continue to apply, while the blocks and actions available in the app determine what the assistant can do.

For example, a Softr builder could use the Softr MCP Server to ask an AI assistant to build a Softr app or make changes to a database, while someone using a published app could use its App MCP Server to look up information or update a record through their AI assistant.

One connector, infinite connections

Before MCP, you always had to juggle between tabs, gathering context in the apps you work with and then dumping it on the AI agent's chat thread. But now, you can chat naturally with your chatbot and it can look up information on its own and start actions for you without the tab-hopping.

If you already built apps on Softr, follow the guide to activate the MCP server in the Settings tab and connect your favorite AI assistant. If you're still running legacy internal tools in platforms that don't support this standard, consider rebuilding them in Softr: you can use the AI Co-Builder to quickly replicate your internal tool lineup, and then activate the MCP server to expose it to your agent.