The selection layer for agents

Agents can call tools. They still need to choose the right one.

ASM turns eligibility, cost, risk, provenance, and policy into a reproducible decision before an agent invokes a service or pays it.

0 / 14,519

Complete selection records in the audited MCP data

Across five registries and directories, no entry combined comparable price, SLA/rate-limit, quality, and access-mechanism signals in machine-actionable form. Runtime payment remains the selected rail's job. Snapshot: May 2, 2026.

6 / 6

Models selected better with structured metadata

In ToolSelect-Bench v0, every tested model improved correct selection with ASM metadata; constraint violations fell for five of six. Raw picks are committed for rescoring.

One concrete decision

Book a flight I can later cancel.

The selector rejects ineligible tools and surfaces financial risk and approval requirements. This uses expired June 2026 example facts, not current provider recommendations.

task: book a flight I can later cancel
taxonomy: tool.booking.travel
requires: flight_search, flight_order_create, booking_cancel
approval gate: financial_charge

Ready. No account or payment required.

The live result will appear here.
The boundary

Discovery → selection → settlement → evidence

1

MCP / ARD / catalogs

What services exist and how an agent reaches them.

2

ASM

Which candidates are eligible, which one fits the user's policy, and why.

3

MPP / x402 / ACP

How authorization and payment execute after a service is selected.

4

Observed evidence

What actually happened, so the next decision can distrust stale or inaccurate claims.