MCP, in plain terms
MCP stands for Model Context Protocol. It is an open standard published by Anthropic in late 2024, designed to connect a language model to external systems (data, tools, actions) without writing a custom integration for every software/model pair. The protocol has become multi-vendor: several AI applications (Claude, ChatGPT, development environments like Cursor or Visual Studio Code) can speak it, and any compliant server can connect to them.
The principle comes down to two roles. An AI application acts as the MCP client: it decides, over the course of a conversation, when to call which tool. An MCP server exposes what it can do in a standardized form: tools, functions the agent can call with defined parameters, and sometimes resources, content that can be read rather than a call. Between the two, the protocol fixes a shared exchange format, so an agent does not have to guess the shape of responses from one server to the next.
The Nautilinks MCP server illustrates both notions concretely: eleven callable tools covering the catalog, existing articles, AI visibility plans, balance, ordering and tracking, plus one resource explaining how the network works.
What it actually changes
Without MCP, an agent that wanted to buy a link would need either a human to do it on its behalf, or an attempt at driving a browser the way a person would, which breaks at the first interface change. With an MCP server, the agent calls the exact same functions that run the site, with defined parameters and a structured response, never touching the page's HTML.
Concretely, on Nautilinks, an agent connected to the MCP server can search the catalog by niche, language or budget, across the three fixed-price shelves (Plancton at 5 €, Corail at 15 €, Nautilus at 30 €), read a site's full record, check the account's prepaid balance, place an order for one to twenty links, then track each order through to publication.
The response to placing an order tells the agent itself how things proceed next: if the account's prepaid balance covers the total, the order is settled immediately. Otherwise, a Stripe payment link comes back, which a human has to open and pay. No card is ever charged without a person approving the payment, either upfront by funding the balance, or at that moment through Stripe.
The server keeps nothing between one order and the next: every call carries the caller's own API key, forwarded as-is to the Nautilinks API, with no session or memory kept on the MCP server side. It is a relay, not a separate account.
Connect an agent
The Nautilinks MCP server is remote (a Cloudflare Worker), over HTTP, with nothing to install. A Nautilinks API key, created in the member area, is sent on every call.
claude mcp add --transport http nautilinks https://mcp.nautilinks.co/mcp \
--header "Authorization: Bearer sn_live_your_key"The detail of all eleven tools, the equivalent REST endpoint table, authentication and sandbox keys (to test without spending) live on the API reference, which documents how to wire it up. This page answers what and why.
Frequently asked questions
Does Nautilinks own MCP?
No. MCP (Model Context Protocol) is an open standard, published by Anthropic in late 2024, that Nautilinks chose to implement for its own catalog. The protocol itself is not tied to any single vendor: several AI applications already speak it, and any other service can expose its own MCP server the same way.
Do I need to install anything to use it?
No. The Nautilinks MCP server is remote, reachable over HTTP, with no package to install and no server to host yourself. An MCP client (Claude Code, claude.ai, or any other tool that speaks the protocol) connects to it directly by URL, with an API key in the header.
Does MCP replace the REST API?
No, both expose the same service. The REST API (/api/v1/agent/*) is built for a developer writing their own calling code. The MCP server exposes the same operations in a form an agent can discover and call on its own, without a hand-written integration. A developer who prefers to call the API directly still can, nothing forces going through MCP.
Who can use my key if I connect it to an agent?
The key is sent on every call, either as a header or a URL parameter depending on the client. The MCP server does not store it, but anyone who obtains it can use it exactly as you would, with the same rights. It is the same trust boundary as calling the API directly: a key can be revoked at any time from the member area.
Can I test it without risking a real spend?
Yes. A sandbox key (sn_test_ prefix) simulates an order without ever charging the prepaid balance or creating a real order with a network publisher.
Can an agent order on its own, with no human involved?
It can search, compare, and place an order on its own. Payment itself never completes without a human step somewhere: either upfront, when a human funds the prepaid balance the agent later spends, or at order time, when a Stripe payment link waits for a person to open and pay it.