Testing an agent's spend guard against a hostile 402 seller
How to test the code that lets an agent pay for API calls against a hostile 402 seller: thirteen cases, an honest control and a mock wallet, offline.
Articles tagged: testing
MCP Safety Policy: product page
Starter allow, confirm and deny rules for the tool calls an agent makes (shell, files, deletion, payments, network), an evaluator, a before-tool-call hook and a token estimate. Offline. MIT licence.
MCP Safety Tests: product page
Offline test fixtures that check whether your MCP client or agent refuses poisoned tools: 33 golden cases in 7 families (tool poisoning, shadowing, rug-pulls and clean controls), a black-box harness and a tool-description linter. Same result every run; no LLM key, no network. MIT licence.