Registry / Methodology

How we score

One rubric, applied to every project, with the math shown. The point is a ranking you can check, not take on faith.

The five criteria

Criterion	Weight	What a 5 looks like
Execution	25%	Order placement, partial fills, and a reliable one-call exit.
Safety	20%	Non-custodial keys, reduce-only, size and loss limits.
Lifecycle	20%	Backtest, paper, deploy, and monitor in one place.
Agent-native	20%	Callable as MCP tools or a typed skill, built for agents.
Openness	15%	Open source, public docs, and an inspectable surface.

How a score is built

Each project gets a 0–5 on every criterion from its public docs, repo, and SDK. We multiply by the weights above and sum to a single 0–5 score, then sort. A tie breaks toward the more open and more agent-native option, because those are easier for you to verify and adopt.

What the scores are, and are not

They are our reading of what each tool publicly supports today. They are not paid placements, and they are not live performance benchmarks. Where we cite a hard number, it links to a source.

Coming next: head-to-head latency and strategy-accuracy benchmarks run on a fixed harness, with the methodology and raw results published here so anyone can reproduce them. Until those are live, the ranking reflects capability on the five criteria, not measured speed.

Corrections

Scores are judgement calls on public information, so some will be off. If an entry is mis-scored or missing, send the repo or docs link with the fix and we will re-score it against the criteria above. No vendor pays to be listed or ranked.