quicksilver
Claude Code skill that hands bulk 'read a lot to decide a little' judgments over files, logs and tickets to Jev and returns a shortlist.
Overview
Quicksilver is a Claude Code skill aimed at the part of a session spent reading material only to decide whether it matters: which of 187 files handle auth, which of 3,000 log lines are real failures, which tickets are refund requests. Instead of Claude reading everything, the skill sends items to Jev in parallel and returns a shortlist with confidence, so Claude spends tokens only on the relevant items.
The README reports a 12-task benchmark (eight on public data such as BGL logs, Banking77, UCI SMS Spam, SST-2, the Hono codebase and lodash) comparing a normal Claude Code subagent against Claude plus Quicksilver. Token use dropped by roughly 76-95% per task and runs were up to 20x faster, at a median of $0.004 of Jev per task and $0.45 for the whole benchmark. Quality was not always equal: on real BGL log triage F1 fell from 54% to 23%, and spam F1 from 97% to 92%. Content is sent to TypeSafe's API.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this file handle auth? (or: is this log line a real failure?) | — |
| Choice | Which intent is this ticket (e.g. refund request)? | — |
| Score | Score items against a criterion for shortlisting(inferred) | — |
How it works
- Claude recognises a 'read a lot to decide a little' task
- The skill sends items to Jev in parallel with a typed question
- Jev returns verdicts with confidence
- Claude receives a shortlist (and borderline items) instead of the full content
Strengths
- README reports 76-95% fewer Claude tokens across 12 tasks
- Up to 20x faster, median $0.004 of Jev per task per the README
- No npm dependencies
Caveats
- Accuracy dropped on some tasks (BGL log F1 54% to 23%)
- Content is sent to TypeSafe's API
- Needs a Jev key
Getting started
npx github:UditAkhourii/quicksilver (asks for your Jev key once), then restart Claude Code.At a glance
- Type
- Agent skill
- Status
- beta
- Language
- JavaScript
- Runtime
- Claude Code skill (Node 18+)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Claude Code skill: hand bulk judgment calls to Jev.
wh000wh000/awesome-jev-liveClaude Code skill for the bulk judgement calls — which of these files, which of these log lines, which of these tickets — that otherwise get read one by one.
MrJev/awesome-jev
Frequently asked questions
What does Quicksilver decide?
Which items in a large set (files, log lines, tickets) match the criterion Claude is looking for.
Does it replace Claude?
No; the README says Jev cannot write, reason or edit, so Claude still does the thinking on the shortlist.
What does it cost?
The README cites $0.042 per million input tokens with free output, and $0.45 for its whole benchmark.
Related projects
CLI, MCP server and Claude Code plugin for closed-form yes/no, choice, score and rank decisions via Jev, OpenAI Decisions or Claude.
MCP serverbetaTypeScriptCommunity MCP server giving Codex, Claude and other clients two tools for typed Jev decisions: evaluate and classify.
MCP serverbetaTypeScriptHermes Agent plugin exposing Jev as four coding-workflow tools: jev_check, jev_route, jev_score and jev_evaluate.
PluginbetaPythonLocal service to define, test and publish versioned Jev judgment functions, callable over HTTP by backends and over MCP by agents.
MCP serverbetaTypeScriptMCP server giving agents typed Jev decisions (ask, choice, score, noul, batch) with calibrated probabilities via the OpenJEV API.
MCP serverbetaTypeScriptTool that lets code and agents ask Jev typed questions via reusable jevels, mapping answers to act, mark or fall_back.
ProjectprototypeTypeScript