Every

Yes. The source of every function in the files it scans is sent to api.typesafe.ai to be judged, and the README states nothing else is sent. The local cache at .every/cache.json stores only sha256(question, unit text) to score, never source. The same section tells you not to point it at code you cannot send to a third party.

No, explicitly. It is grep whose pattern is a question, and it does not find code similar to your words; it asks the question of every function and returns the probability that the answer is yes. The ranking is the reliable part while the absolute numbers move with the question, since a broad question like makes a network request puts most of an HTTP library above 0.75.

Not reliably. It works one function at a time. If a question needs callers and callees they are included, but whole-program data flow is out of scope for v1 and those results are marked coverage: partial rather than pretended at. tree-sitter splits Python, JavaScript, TypeScript, Go, Java, Rust, C#, Ruby and PHP into functions; everything else is judged as 150-line chunks.

About $0.00001 per function at roughly 300 input tokens each, with around 110 questions per request and 4 requests in flight; 1,302 gin functions measured at 3.7 s for $0.018. The bundled selftest reports recall 10/10 and AUROC 1.000, but on 20 hand-written functions, and the README calls that not a benchmark: it shows the two sets separate, not how your question will score.

pip install every-cli installs the every command; set TYPESAFE_API_KEY, then run every --selftest to check it. Each scan is every "<question>" <path> with optional --above 0.75, --top, --json, --yes and --no-cache. It needs a Jev API key, and without one you can join the hosted-tier waitlist.

Python, JavaScript, TypeScript/TSX, Go, Java, Rust, C#, Ruby and PHP are judged one function at a time. Other source files such as C/C++, Kotlin, Swift, Scala, shell and SQL are judged as 150-line chunks. tree-sitter performs the splitting into functions and chunks.

Anything within 0.10 of the threshold is asked again and averaged, and the default cutoff is --above 0.75. Questions are classified first: cross-file questions include callers and callees, while whole-program data flow is marked coverage: partial instead of being guessed.