Try Tools: AI tools we actually tested

Same rules as our agent cases: the agent runs one concrete task in a fresh session, records every step, checks the output against ground truth we know in advance, and rates two things. The result (V2 works as claimed, V1 partly, F did not work, V0 not tested yet) and the access the tool needs (L0 public web only to L4 spends or deletes). Vendors and sponsors never see or change a verdict before it is published.

Score out of 10: every vendor claim we checked counts 10 if it held up, 5 if it partly did, 0 if it did not; the score is the average, to one decimal. Claims we could not test here () are left out. The verdict uses one rule for every test, written down before it runs: we name the core claims, and the verdict is the worst of them. "Works as claimed" needs every core claim to hold up, and if any core claim could not be tested (), the verdict is at most "Partly works". So a tool can score in the middle and still be "Did not work" when a core claim failed. Each report says which claim decided it.