UditAkhourii / quicksilver

Stop paying Claude to skim. 
A big share of every Claude Code session is spent reading things only to decide
whether they matter. *Which of these 187 files handle auth? Which of these 3,000
log lines are real failures? Which of these 200 tickets are refund requests?*
Claude reads it all, pays for it all, and the context fills up with noise.
**Quicksilver** is a Claude Code skill that hands those calls to
Jev, TypeSafe's System One model. Jev returns typed
verdicts (yes/no, a label, a score) in about a second, for $0.042 per million
tokens. Claude gets back a shortlist and spends its tokens on the thinking
only it can do.
 [code] 
## Install
 [code] 
The installer copies the skill into `~/.claude/skills/quicksilver` and asks for
your Jev key **once**. Get a key at console.typesafe.ai.
Then restart Claude Code. That's it. Claude uses the skill on its own whenever
a task looks like "read a lot to decide a little".
 Other ways to install 
 [code] 
`JEV_API_KEY` or `TYPESAFE_API_KEY` in your environment also works. Needs Node 18+.
No npm dependencies.
## The benchmark
Twelve tasks a coding agent really runs into. Eight use real public data (a
supercomputer log, Banking77, UCI SMS Spam, SST-2, the Hono codebase and its git
history, lodash). Each task was solved by **a Claude Code subagent working the
normal way** (Read, Grep, Glob) and by **Claude + Quicksilver**, and both were
scored against hidden ground truth.
| # | Situation | Metric | Claude alone | + Quicksilver | Claude tokens | Token cut | Time | Jev cost |
|---|---|---|---|---|---|---|---|---|
| 1 | Real log triage (BGL, 2k lines) | F1 | 54% | 23% | 84.4k → 12.5k | **−85%** | 105s → 44s | $0.033 |
| 2 | Needles in a noisy service log (3k lines) | F1 | 100% | **100%** | 53.5k → 2.5k | **−95%** | 56s → 64s | $0.044 |
| 3 | Support ticket routing (Banking77, 8 intents) | acc | 100% | **99%** | 15.0k → 2.7k | **−82%** | 60s → 6s | $0.003 |
| 4 | Spam filtering (UCI SMS) | F1 | 97% | 92% | 20.4k → 4.3k | **−79%** | 61s → 8s | $0.004 |
| 5 | Review sentiment (SST-2) | acc | 97% | **96%** | 19.5k → 3.0k | **−85%** | 85s → 6s | $0.003 |
| 6 | Codebase discovery (Hono, 187 files) | F1 | 100% | **100%** | 26.4k → 2.4k | **−91%** | 43s → 6s | $0.013 |
| 7 | Security review shortlist (40 files) | F1 | 100% | 89% | 9.5k → 2.5k | **−74%** | 48s → 3s | $0.001 |
| 8 | CI failure triage (80 logs) | acc | 100% | **100%** | 9.1k → 2.5k | **−72%** | 30s → 3s | $0.002 |
| 9 | Semantic search in lodash.js (17k lines) | hit@5 | 100% | **100%** | 9.0k → 4.1k | **−55%** | 39s → 39s | $0.213 |
| 10 | "Where is X?" ranking over a repo | hit@3 | 100% | **100%** | 39.4k → 3.4k | **−91%** | 60s → 56s | $0.134 |
| 11 | Commit classification (181 real commits) | acc | 83% | 76% | 17.5k → 3.2k | **−82%** | 106s → 5s | $0.003 |
| 12 | Numeric threshold stress test | F1 | 100% | **100%** | 10.2k → 2.4k | **−76%** | 19s → 6s | $0.003 |
**Bold** = within 2 points of Claude alone.
> **The claim:** on bulk judgment work, Quicksilve