MCP Leaderboard

The INITE Atlas asks every vendor's own MCP server once a day whether it answers, openly or asking for a sign-in, and how fast; an open server's tools are read and graded. The score is the Atlas's own, out of 100: reliability over 30 days, the share of the product's use cases the tools do, how well the tools are defined, what they cost an agent's context, public adoption and how readily a client gets in. Its weights are on the Atlas method page.

The first board is the servers people know: the Atlas's most used vendors whose server answered at least once this week, in order of use, with their score where they have one. Most of them sit behind a sign-in, so they have no full record and are not on the score board.

The board ranks only servers with a full record: at least a week of checks, tools read, use cases judged. A server behind a sign-in shows nobody its tools, so it is counted in the week's numbers but not ranked. Ties are broken by the week's checks, then by how much the vendor is used in the Atlas.

A mover's score is compared with the score frozen the week before, and a change under 10 points is left out; its uptime is compared with the week before's checks, and a change under 30 points is left out. A server is listed as not answering our checks when fewer than half of at least four checks got an answer after it had been up. That is what our checker saw, not a verdict on the vendor: a firewall, a regional block or a rate limit on our side looks the same. Nothing here is written by a model: the sentences are filled in from the numbers in the tables.

Free to cite with a link to this page (CC BY 4.0).