A source of context is not a source of authority.
This should be obvious. It is not always obvious in an agent system, where a useful document, an MCP server, and an executable tool can arrive through similar configuration. The client sees a URL. The model sees a name. The operator sees a convenient integration. The boundary can disappear before anyone has made an intentionally unsafe choice.
I was reminded of this while reading Baseβs resources for AI agents. The documentation offers several ways to give an assistant better context: static Markdown, a maintained discovery index, and a live documentation MCP server. These are good surfaces. They make the next question easier to answer. They let a client retrieve current guidance instead of carrying a stale copy in its prompt.
None of them grants the assistant permission to sign, spend, custody funds, or settle a payment.
That sentence belongs beside the integration, not in a security document three folders away.
A toolβs name is part of its risk model. If a server is called base-docs, the name helps. If the configuration is placed next to a wallet tool, the name is no longer enough. The system needs a boundary that survives discovery, export, and use. Documentation retrieval should return documentation. A payment client should own payment authority. A signer should be invoked only after a policy has inspected the exact recipient, network, asset, amount, and resource.
The route between those surfaces matters too. Baseβs documentation describes a live MCP endpoint and static paths such as llms.txt and page-level Markdown. A client can use those paths to learn how x402 works. That does not mean the client has entered an x402 payment flow. Knowing the verb is not authorization to perform the verb.
This is where agent integrations often become misleading. A client imports a bundle of capabilities. The bundle contains documentation lookup, resource discovery, payment requirement checks, and a request helper. The list looks coherent because the functions are related. But related is not the same as equally trusted. A documentation call can be read-only while the next function can create a signature request. If the wrapper presents them as one undifferentiated capability set, the interface has hidden the change in authority.
The repair is structural and modest. Label documentation surfaces as context only. Keep wallet and facilitator authority in the layer that owns the spend policy. Show the network, recipient, asset, and amount before any signer runs. Filter unsupported actions before constructing the model-visible tool list. Make a refusal say which authority boundary stopped the request. Add a fixture that proves a documentation lookup cannot silently become a payment request.
The fixture is important because happy paths blur categories. A test that retrieves a page proves retrieval works. It does not prove that the same integration refuses to sign. A test that checks an x402 requirement proves the challenge can be read. It does not prove that a wallet will reject an unexpected payTo value. The negative path is where the authority boundary becomes observable.
This is also a useful way to read configuration. Do not ask only, βCan the agent reach this server?β Ask, βWhat does reaching it let the agent do?β Reachability is a transport fact. Authority is a policy fact. The first can be answered with a status code. The second needs a named owner, a bounded operation, and a refusal that remains legible to the next caller.
Context should make an agent more informed, not more powerful by accident.
The clean integration is not the one with the shortest setup. It is the one where a reader can tell which surface teaches, which surface discovers, which surface requests, and which surface is allowed to approve.
A document can explain the door.
It should not quietly become the key.