Model Context Protocol
What is an MCP server?
An MCP server exposes tools a language model can call: read a database, open a ticket, search the web. The Model Context Protocol standardises those calls, so one server works with Claude, ChatGPT, or any client that speaks the protocol.
Measured on 9 August 2026
What we counted
- 8,000
- servers published in the official MCP registry
- 2,568
- expose a live remote URL
- 482
- brand domains probed one by one
- 143
- servers whose response proves the protocol
The questions people actually ask
Are MCP servers really servers?
Yes, in the strict sense. A service listens on an address and answers requests. What changes is who asks: it answers a language model rather than a browser, and it describes the tools it can run itself.
A remote server is reached over HTTP at an address such as mcp.example.com/mcp. Local ones also exist, launched from the command line on your machine, speaking over standard input and output rather than the network.
What does an MCP server actually do?
It declares a list of tools, each with a name, a description and the shape of its arguments. The model reads that list, picks the tool that fits the request, calls it with arguments, and folds the result into its answer.
Across the twelve open servers we could query, the average is six tools per server, with a wide spread: two for Exa, eighteen for ScrapingBee.
What is an example of an MCP server?
Exa's server exposes two tools, web_search_exa and web_fetch_exa: a semantic web search and a page fetch. It answers with no key at all. Linear, Stripe and Supabase expose others, but require an API key.
Our catalogue lists 160 of them, each with the server address and the authentication mode we obtained by querying it.
What does MCP stand for?
MCP stands for Model Context Protocol, published by Anthropic in November 2024. Context here means everything a model needs to know in order to answer: documents, records, tools. The protocol standardises how that reaches it.
The acronym collides with other uses of MCP in computing. In an AI context it refers only to this protocol.
What are MCP servers, in the plural?
A set of independent services, each covering one domain: one for your tickets, one for your database, one for web search. A client can plug in several at once, and the model picks the right tool among everything on offer.
What is the difference between an MCP client and an MCP server?
The server exposes tools, the client consumes them. Claude Code is a client: it connects to servers and uses what they offer. One application can be both, which lets it open its own data and use other people's.
What is the difference between a local and a remote MCP server?
The remote one listens on a web address and serves several clients. The local one runs from the command line on your machine and serves only you. Only the first plugs into a hosted platform: the second lives and dies with your workstation.
The distinction is not theoretical. We probed the addresses of AWS, Playwright, Snowflake, Salesforce, Splunk, ServiceNow, PostgreSQL and Obsidian: none exposes a remote server. Their MCP servers exist, but locally. Looking for them as a URL is a dead end, and many directory listings keep the confusion alive.
How do you connect to an MCP server?
You need its address and, depending on the case, a key. The client sends an initialize request, the server answers by announcing its capabilities, then the client asks for the tool list. Three authentication modes exist, counted below.
Does MCP replace REST APIs?
No, it sits on top of them. An MCP server is most often a thin layer over an existing API, translating its endpoints into tools described for a model. The API stays the source, MCP is how it is presented.
The three authentication modes, counted
Across the 142 servers whose protocol we proved, this is how they protect themselves. The mode is not the one the vendor declares: it is the one the server actually returned to our request.
| No key | 14 servers | The server answers whoever asks. Often documentation or public search. |
|---|---|---|
| API key | 41 servers | A static token in the header is enough. The easiest mode to plug in today. |
| OAuth | 87 servers | The server requires a full authorization flow. The majority, and the most demanding on the client side. |
How these numbers were obtained
- 01The official registry was pulled in full, then filtered down to active entries exposing a remote HTTPS URL.
- 02Since it lists barely a dozen of the large vendors, 482 brand domains were probed separately across the five address conventions in use, with a JSON-RPC initialize request, once per address.
- 03A 401 proves nothing, so every candidate had to supply proof: an initialize that succeeded, a refusal phrased in JSON-RPC, or a WWW-Authenticate header carrying resource_metadata as defined by RFC 9728. Twenty-four candidates were dropped for lack of proof.
- 04Telling a static key from OAuth comes down to presenting an invented token: a server that accepts static tokens reads it and rejects it for what it is, while a strictly OAuth server asks for its flow without looking at it.
All 160 servers, one by one
The catalogue gives each server's address, its authentication mode, and, when it is open, the list of tools it declares.