Observed arrival · 2026-11-21
Toolmark records what AI-agent tools do in a sandbox
It runs exact versions of MCP servers, skills, and plugins in an isolated sandbox, then compares observed behavior with the tools’ declared claims.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Teams reviewing MCP servers, agent skills, plugins, and API documents
- Worth noticing
- Outbound-capable tools are called last, only after network closure is verified, and with synthetic values.
Field notes
A package check starts by resolving the requested exact version, then runs in a fresh cloud sandbox with no network, no secrets, and a synthetic data folder. The sample contrasts declarations such as a “read-only” hint with observed changes, and marks uncalled tools as “not tested” rather than treating them as failures. The displayed example is described as a run against Toolmark’s own server with planted flaws.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue