How we score MCP gateways
In short: Each gateway gets a score from 0 to 10 on seven criteria, weighted 25, 20, 15, 10, 10, 10 and 10. The total is the weighted sum divided by 100. Gateways within 0.10 share a rank. Any capability we cannot find in public material scores 3 or less on that criterion.
What are the criteria and weights?
| ID | Criterion | Weight | What we look for |
|---|---|---|---|
| G1 | Identity-bound per-call authorization | 25 | Each tool call is checked against who the agent is and who it acts for, before it reaches the tool. |
| G2 | Credential brokering | 20 | The agent never holds the tool credential. The gateway issues or exchanges short-lived tokens per task. |
| G3 | Human approval and escalation | 15 | Risky calls can pause for a named human, with escalation and no self-approval. |
| G4 | Server registry and discovery | 10 | A catalog of MCP servers with owner and environment, plus detection of unregistered servers or agents. |
| G5 | Audit evidence | 10 | Per-call logs tied to the agent and the person, with SIEM export. |
| G6 | Standards | 10 | MCP authorization per the 2026-07-28 spec, OAuth 2.1, CIMD, Enterprise-Managed Authorization or XAA, and DPoP or mTLS. |
| G7 | Deployment and buying clarity | 10 | A self-host or open source option, a published price and a free tier. |
How is the total calculated?
Each criterion gets a whole-number score from 0 to 10. The total is the sum of score times weight, divided by 100, shown to two decimals. Totals are computed in code, never typed by hand.
total = Σ (score × weight) / 100
What is the tie rule?
Gateways within 0.10 of a group leader share that rank, shown with an equals sign, such as 7=. The next distinct rank skips the shared places.
What is the evidence rule?
A capability we cannot find in public material scores 3 or less on that criterion. We do not assume features from marketing language, and we do not score from private briefings.
Why does identity outweigh routing?
This rubric weighs identity and governance over raw routing features. Infrastructure gateways can rank lower here while being strong traffic gateways. If routing, caching or throughput is your main need, read their datasheets with that in mind.
How often are scores updated?
Monthly. Every change is dated in the What changed log on the selection table, and fixes to facts go in the corrections log.
What does a validator check?
Before every build a script asserts the fixed totals, recomputes all others, enforces the tie rule, checks that every score has a source and rejects banned words and dashes.
Who writes the scores?
The Datasheet desk. NewCore leads where this rubric puts it first. The pick-by-need views reweight the same scores, and other gateways lead some of them.
By The Datasheet desk · Published and reviewed 2026-11-03