Picture this. You run the backend for a product that serves many customers, and you want an AI agent to work inside Notion for one of them. You have an MCP server in front of Notion. It knows which account it authenticated. What it does not know is which of your customers the agent is supposed to be working for. That gap is the one most MCP tutorials skip, and it is the one you would be answering for if the agent read the wrong customer's pages.
Nango's guide on how to add access control to any MCP server takes that question head on. Its answer is Agent Sessions. Your backend verifies the user's tenant membership first. Then it creates a session that selects that tenant's connections by tag, allowlists named tools, and expires on a fixed clock. In the walkthrough, CustomerA's agent gets exactly two Notion tools, notion-search and notion-fetch, for one hour. Nothing changes on the upstream server.
You would notice where the decision sits. Your backend makes the tenant call before the agent starts, and the agent, in the post's words, "cannot change the scope of its session." You are not asking the model to behave. You are handing it a session that is already narrow.
The guide is also blunt about its own product, and you should take the warning seriously. Keep nango_proxy disabled, because it "bypasses named-tool allowlists, including direct calls to upstream MCP tools." If you switch it on, the two-tool allowlist you just built no longer describes what the agent can do.
There are limits you would carry into your own design. Nango says plainly that if one connected Notion account can see both customers' pages, tenant tags will not separate them. Connection selection is not resource permission. And if you bring your own MCP server, it must support OAuth with Dynamic Client Registration or Client ID Metadata Documents.
The figures in the post are Nango's: "10,000+ prebuilt tools across 1,000+ APIs." Treat them as vendor counts. The post's view is that what survives without them is the design, and that the design is a good one.
Now the turn. The APIs.io catalog holds a record on Nango, and that record is where you would go to check how an agent could read all this. The provider page lists six API pages, and the story walks four of them. Tenant-tagged connections are the Nango Connections API. The Notion MCP setup lives in the Nango Integrations API. The prebuilt tools exposed through a session are the Nango Actions API. The raw passthrough you were told to switch off is the Nango Proxy API.
What the catalog does not have is the session endpoint itself, the thing the whole guide is about. The agentic access profile counts 21 operations, 14 of them acting, and the catalog description still says "over 400 APIs with 600+ pre-built integrations." The product in the post is newer than the record that describes it.
The scores point the same way. Nango scores 52.6 on the Kin Score, which the catalog calls developing. Access clarity at 71.1 and discoverability at 66.7 carry it, while contract governance sits at 13.6. Agent Readiness is 31.6, and the lit dimensions match the post: an MCP server, auth clarity, protected resource metadata.
The dark ones are the instructive part, and you would want to know them before you build. Dynamic client registration is dark on the company whose generic MCP path requires it of everyone else. Error semantics and idempotency are dark on a platform whose whole pitch is that an agent's calls are executed safely on someone's behalf. The post's conclusion is that the access-control model is ahead of the contract that would let an agent read it.
So here is what you take away. The pattern is worth copying: decide the tenant in your backend, give the session a short clock and a named-tool list, and leave the proxy off. Check one more thing Nango itself flags, which is whether the upstream account can see more than one customer, because tags will not fix that. And when you read Nango's catalog record, read it as a record that trails the product. The guide is the better source for how sessions work today.
Read the original: https://apis.io/2026/11/16/nango-scopes-mcp-agents-by-tenant-and-its-catalog-record-is-a-product-behind/