← Back to the overview

Assessment methodology — agent access

This page describes rubric version v1

Scores are calculated from publicly documented facts, not tested ourselves. Every exchange receives two scores on a scale of 0 to 10: an agent score (how mature the agent layer is) and a safety score (how limited the damage an agent can cause is). A field we could not verify does not count as a poor result — it simply does not count.

Agent score (10 points)

Coverage (0–4): official write access, read-only access, or community tooling, plus bonus points for the number of tools and the breadth of supported asset classes. Interface (0–2): structured JSON output and a stable error taxonomy. Sandbox (0–2): a test environment and dry-run mode. Distribution (0–2): ready-made skills and client independence (any MCP client, not a fixed list).

Safety score (10 points)

Impact limitation (0–4): a segregated agent account, a spending limit, and a separately revocable withdrawal right. Approval (0–2): whether per-order confirmation is required, optional, or absent. Key management (0–2): whether the model itself can access the key, and whether a narrowly scoped key (buy-only, spot-only) is available. Recoverability (0–2): a dead man's switch and notifications on every trade. A read-only agent layer cannot place orders or move funds and therefore always scores at least an 8 on safety.

How unknown fields are handled

An unknown field earns no points, but also does not count toward the denominator the score is measured against. An exchange we have not yet been able to research thoroughly therefore never scores worse than an exchange that has been researched and scores poorly. We only show the score once at least 70% of the relevant fields are known; between 40–70% we show the score with a 'partially assessed' marker; below that we show no score.

Inclusion criteria

Tier 1 includes providers with an official agent layer that are MiCA-licensed, have a MiCA application pending, or are included as an architectural reference point.

  • Binance is excluded: reported as absent from the ESMA register, application withdrawn, EU exit as of July 2026. Binance's agent layer shows nothing the other providers don't already demonstrate.
  • Robinhood is included and marked as unavailable to EU customers. It is the only example of a segregated agent account — the safety benchmark every other provider is measured against — and earns its place as a reference architecture, not as an option for the reader.

Tier 2 includes providers with a public API but no official agent layer, with a material presence in a market we serve.

Tier 3 includes smaller providers or providers active in a single market, with no public API and no agent layer. Listed in the table for completeness, not scored.