Question 1

Does it match Claude's accuracy, or only save tokens?

Accepted Answer

Both, unevenly. The twelve-task benchmark reports an 86% token cut with a median of 82% and parity on eight of twelve tasks, but it also records regressions: log triage F1 falls from 54% to 23%, the security shortlist from 100% to 89%, and commit-type labels from 83% to 76%. Treat those three as shortlist-only cases.

Question 2

What should I not send through it?

Accepted Answer

Anything whose labels encode policy nobody wrote down. The README's own worst case is the BGL supercomputer log, where alert marks many FATAL lines as normal — Claude alone reached 54% F1 there and Quicksilver 23%. It also flags look-alike code at borderline probability and subjective house-style labels such as perf versus refactor as cases where Claude should check the ? items.

Question 3

Is it faster than Claude on large scans?

Accepted Answer

Not always. On huge scans involving thousands of Jev calls, wall-clock time is about the same as Claude's — the win there is tokens and context, not speed. Where it does win, the README reports roughly 10x faster on bulk routing, and `--fast` packs items to go about 10x faster on obvious needles. It is explicitly not for writing, editing, multi-step reasoning, or anything grep answers exactly.

Question 4

What actually leaves my machine?

Accepted Answer

File content goes to TypeSafe's API at api.typesafe.ai, so the README says not to point it at anything you cannot send to a third party. It never sends .env files, private keys, certificates or credentials, respects .gitignore, and skips binaries and files over 2 MB. Your key lives in ~/.quicksilver/config.json with user-only permissions, and `setup --remove` deletes it.

Question 5

How do I install it, and what does it need?

Accepted Answer

One command: npx github:UditAkhourii/quicksilver. The installer copies the skill into ~/.claude/skills/quicksilver and asks for your Jev key once; get a key at console.typesafe.ai. It needs Claude Code and Node 18 or newer, has no npm dependencies, and JEV_API_KEY or TYPESAFE_API_KEY in the environment also works. Restart Claude Code afterward.

Question 6

What commands does it add, and what does the output look like?

Accepted Answer

Commands cover qs filter for file screening, qs classify for bulk labels, qs rank for relevance, qs find for lines in huge files, plus qs ask and qs status. Output is built for an LLM to read: one line per hit, repeated log patterns collapsed into line-number ranges, id lists for classify, and a ? on anything borderline.

Question 7

What does Jev cost here, and can I see how much I have saved?

Accepted Answer

Jev is billed at $0.042 per million input tokens with output free, and the whole twelve-task benchmark cost $0.45 of Jev. Every run ends with a receipt of items scanned, matches, borderline count, elapsed time, Jev tokens and cost, and Claude tokens avoided; qs status shows a key check plus lifetime tokens saved.

Quicksilver

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community