What is MCP? The open protocol behind clariBI's integrations
MCP, the Model Context Protocol, is an open standard for connecting AI applications to other software. What it is, how it works in plain terms, what changes for the people using it, where its limits are, and the two ways clariBI uses it.
If you have seen "MCP" in product announcements and wondered whether it matters, the short answer is yes, if you want an AI tool to work with the software you already use. This post explains what the Model Context Protocol is, how it works without the jargon, where it falls short, and how clariBI uses it.
MCP in one paragraph
MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic in November 2024 and now supported by many AI applications and software vendors, that defines how an AI application talks to another system: how it finds out what the system can do, how it asks for something, what comes back, and how access is authorized. The usual comparison is a common plug. Instead of every AI product writing a custom connector for every app, an app offers one MCP server and any MCP-capable AI product can use it.
The problem it solves
Before MCP, an AI tool that wanted to read your billing data had to learn that vendor's API: its authentication, pagination, rate limits and error codes. Then it had to do the same for the CRM, the project tracker and the analytics tool, and keep every connector working as each API changed.
With ten AI tools and a hundred apps, that is up to a thousand separate connectors. With a shared protocol, each AI tool implements MCP once and each app offers one MCP server: a hundred and ten pieces instead of a thousand.
How it works, in plain terms
There are three roles:
- Host: the AI application you use, such as a chat assistant, a code editor or a BI tool.
- Client: the part of the host that holds a connection to one server.
- Server: the program that exposes an app's capabilities. The app's vendor usually runs it.
A server offers three kinds of things:
- Tools: actions the AI can call, each with a name, a description and a schema for its inputs. "List payments between two dates" is a tool.
- Resources: data the AI can read, such as a file or a record.
- Prompts: ready-made instructions the server suggests for common tasks.
When a client connects, it asks the server for its list of tools. Because every tool describes its own inputs, the AI can work out which one fits a request and how to call it, without anyone writing code for that app. Messages use JSON-RPC, a simple request and response format. Local servers can run on your own machine; remote servers are reached over HTTP. For remote servers, the protocol builds authorization on OAuth 2.1, the same kind of sign-in-and-approve flow you know from "Sign in with Google". Many servers also support dynamic client registration, so an AI tool can register itself without a manual setup step; the post on OAuth 2.1 with dynamic client registration explains it.
What changes for the people using it
- Connecting an app is usually a sign-in. You approve access on the vendor's own page, and you can revoke it there later.
- New apps arrive faster. Once a vendor publishes an MCP server, any MCP-capable tool can connect to it without waiting for a custom integration.
- Answers can use current data. An AI tool can call an app's read tools when you ask a question, not only rely on the last sync.
Where MCP falls short
It is a young standard, and some limits are worth knowing:
- Coverage depends on the vendor. A server exposes what its vendor chose to expose. Some offer dozens of tools, some a handful. Shopify's Storefront server, for example, reads the public product catalog but not orders or revenue.
- Tools can write, not only read. Many servers include tools that create, update or delete things. An AI tool connecting to them needs rules about which tools it may call.
- Tool results are text an AI reads. A record containing crafted instructions could try to steer the model. Careful clients treat tool results as data, not instructions, and limit what the model can do with them.
- Servers change. Vendors add, rename and remove tools, so a connection that worked last month may return different data today.
How clariBI uses MCP: two directions
1. Connecting your apps to clariBI
clariBI acts as an MCP client. Its catalog lists 86 apps connected through MCP, including Stripe, HubSpot, PostHog, Notion, Linear, Intercom, Mixpanel, PayPal and Square. Of those, 52 are available in the Trial and on every plan from Starter up, 32 from Starter up, and 2 from Professional up. MCP apps are not part of the Free or Lite plans.
What happens when you connect one:
- Sign in. Most apps use OAuth; a few use an API key. Credentials are encrypted before they are stored.
- Choose what to sync. clariBI reads the app's list of tools and their input schemas, and picks the ones that return data worth keeping, such as payments, deals or events, filling in date ranges for you.
- Use the data. Synced rows get an automatic dashboard like any other source. When you ask a question, clariBI can also call the app's read tools directly for the latest figures.
Only read tools are used; tools that create, update or delete anything are filtered out. Each app in the catalog was checked by us before it was listed, while the vendor runs the server itself. On Starter and above you can also add your own MCP server by URL, with the same read-only rule. The security model post goes into the details, and why we bet on MCP covers the reasoning.
2. Using clariBI from your AI assistant
clariBI also runs its own MCP server. Connect Claude, Cursor or another MCP client to it and you can list your dashboards and reports, look at a data source's columns, run an analysis or generate a report from the chat. It offers 26 tools, and analyses, reports and forecasts use AI credits from your plan as they do in the app. It is available on the Trial, Starter, Professional and Enterprise plans. Clients sign in with OAuth 2.1, or with a key that starts with claribi_mcp_; the authentication help article explains the difference between app sign-in, API keys and MCP keys.
A next step: in a trial, connect the one MCP app your weekly numbers depend on, often Stripe or HubSpot, open the dashboard clariBI builds from the synced data, and check one figure against the app itself. That tells you more about MCP in practice than any explanation.