Quicksilver
Claude Code skill that hands bulk filter, classify, rank, find, and ask calls to TypeSafe Jev, with a 12-task README bench claiming 86% fewer Claude tokens.
Overview
Quicksilver (github.com/UditAkhourii/quicksilver, MIT) is a Claude Code skill that hands those bulk judgments to TypeSafe Jev. Jev returns a yes/no, a label, or a score. Claude gets a shortlist and spends tokens on the thinking only it can do. Install is 1 command: npx github:UditAkhourii/quicksilver. The installer copies the skill into ~/.claude/skills/quicksilver and asks for a Jev key once, from console.typesafe.ai. Restart Claude Code. The same skill installs as a plugin, or from a clone with install.sh or install.ps1. Pass the key with install --key, or set JEV_API_KEY or TYPESAFE_API_KEY. Node 18+. No npm dependencies. The qs CLI is filter, classify, rank, find, ask, and status. The README's 12-task bench says Quicksilver cuts the tokens Claude spends by 86% (median 82%), matches Claude on 8 of 12 tasks, and runs up to 20x faster, at a median of $0.004 of Jev per task. Public counts were 98 stars and 5 forks when this listing was drafted.
Problem: A large share of a Claude Code session is spent reading only to decide whether something matters: which of 187 files handle auth, which of 3,000 log lines are real failures, which of 200 tickets are refunds. Claude reads it, pays for it, and the context fills with noise.
Built for: Claude Code users who want bulk filter, classify, rank, find, and ask calls handed to TypeSafe Jev, and who will read the README misses before they trust a shortlist.
First indexed on Jev Directory: 2026-10-03
Creator and team
- Name
- Udit Akhouri
- Handle
- @UditAkhourii
“Stop paying Claude to skim.”
How Jev is used
- Role in the product flow
- Claude Code skill and qs CLI: bulk filter, classify, rank, find, and ask calls go to TypeSafe Jev, and Claude reads the shortlist
- Primitives
- ChoiceScoreNoul
- State in
- One narrow question plus the files, log lines, tickets, or a state file. filter is a yes/no over paths or lines. classify takes a closed label list. rank and find take a relevance question. ask takes one question and a --state file.
- Decision out
- One line per hit, written for an LLM to read. Repeated log patterns collapse into line-number ranges. Classify results come back as id lists. A ? marks a borderline item. Every run ends with a receipt: scanned, matched, borderline, seconds, Jev tokens and dollars, and Claude tokens not read.
- npx github:UditAkhourii/quicksilver, or install --key, or the Claude Code plugin marketplace command, or install.sh from a clone. Node 18+, no npm dependencies
- Provide a key once from console.typesafe.ai, or export JEV_API_KEY or TYPESAFE_API_KEY. The file copy lives in ~/.quicksilver/config.json with user-only permissions
- Restart Claude Code. The skill runs when a task looks like reading a lot to decide a little
- qs filter, qs classify --labels, qs rank --top, qs find, or qs ask --state. --fast packs items when the needles are obvious
- Read the shortlist and the receipt. Review ? rows in Claude. qs status checks the key and lifetime tokens saved
Quicksilver is the Claude Code skill and the qs CLI. dbreunig-building-with-jev-skill is Drew Breunig's SKILL.md for Choice, Score, and Noul question design on jev-1.13. It does not ship qs or a 12-task bench. kerpopule-hermes-jev-skills is a Hermes skill drawer for routing, memory, compaction, and GUI steps. dzhng-jevgrep walks a repository and returns excerpts for the agent to read. tamaratran-fast-jev-compaction scores tool-result rows so a transcript can drop junk. Quicksilver spends Jev on the bulk call before Claude reads: filter, classify, rank, find, and ask. The README's own limits stay in the page. BGL alert labels fell from 54% F1 to 23%. Security look-alikes are marked ?. Commit types scored 76% against Claude's 83%. Huge scans can tie Claude on the clock and still cut tokens. Cite github.com/UditAkhourii/quicksilver README and bench/README.md.
Sourced performance claims
- README claim on the 12-task bench: Quicksilver cuts the tokens Claude spends by 86% (median 82%). It matches Claude's accuracy on 8 of 12 tasks (bold in the table means within 2 points) and runs up to 20x faster, for a median of $0.004 of Jev per task. The whole bench cost $0.45 of Jev. Jev is $0.042 per million input tokens, and output is free.Source: github.com/UditAkhourii/quicksilver README The benchmark and FAQ
- Table rows the README marks as matches: noisy 3,000-line log F1 100% to 100%, Claude tokens 53.5k to 2.5k (-95%), time 56s to 64s, Jev $0.044. Banking77, 8 intents, accuracy 100% to 99%, 15.0k to 2.7k (-82%), 60s to 6s, $0.003. Hono codebase discovery, 187 files, F1 100% to 100%, 26.4k to 2.4k (-91%), 43s to 6s, $0.013. CI failure triage, 80 logs, accuracy 100% to 100%, 9.1k to 2.5k (-72%), 30s to 3s, $0.002. Commit classification, 181 real commits, went from 106s to 5s at accuracy 83% to 76%.Source: github.com/UditAkhourii/quicksilver README benchmark table
- README misses: BGL supercomputer log, 2k lines, F1 54% to 23%, tokens 84.4k to 12.5k (-85%), 105s to 44s, $0.033. Security review of 40 files, F1 100% to 89%, tokens -74%. Spam filtering F1 97% to 92%. Commit types 83% to 76%. The README says look-alike safe files landed near p 0.5 to 0.65 and are the ? rows. On huge scans, wall-clock time is about the same as Claude. The win there is tokens and context.Source: github.com/UditAkhourii/quicksilver README benchmark table and Where it shines
- bench/README.md token rules: Claude-alone totals subtract a 68.5k-token fixed floor, measured with a control agent that reads 1 tiny file and writes 1 line. Quicksilver is charged the full SKILL.md on every situation, plus each command, its full stdout, and 40 wrapper tokens per call. Counts use characters divided by 4. Claude-alone agents used Sonnet. 8 tasks use public data. Synthetic sets S02, S07, S08, and S12 were written by the author. S12 is a numeric negative control. The main table still shows F1 100% to 100% and -76% tokens on that row, and bench/README.md says plain awk wins there.Source: github.com/UditAkhourii/quicksilver bench/README.md
- Safety: the skill never sends .env files, private keys, certificates, or credential files. It respects .gitignore and skips binaries and files over 2 MB. Content goes to api.typesafe.ai. The README says TypeSafe states that Jev is not trained on customer data. npx github:UditAkhourii/quicksilver setup --remove deletes the local key. The project says it is independent and not affiliated with Anthropic or TypeSafe AI.Source: github.com/UditAkhourii/quicksilver README Safety and FAQ
- Public GitHub repo UditAkhourii/quicksilver had 98 stars and 5 forks on 2026-10-03. License MIT. Language JavaScript. Created 2026-09-25. Last push 2026-09-25. README copyright line names Udit Akhouri.Source: GitHub API 2026-10-03 and the Quicksilver README license line
Features and stack
Features
- 1-command npx install into ~/.claude/skills/quicksilver, plus a Claude Code plugin path
- qs filter, classify, rank, find, ask, and status
- Shortlist output with collapsed repeats, id lists, and a ? on borderline rows
- 12-task bench with a reproducible bench/ directory and a published token-counting rule
- Skip rules for secrets, gitignored paths, binaries, and files over 2 MB
- Receipt on every run: Jev dollars and Claude tokens not read
Stack
- JavaScript
- Node 18+
- Claude Code skill
- TypeSafe Jev
- MIT
Pricing: MIT code. Jev calls spend your TypeSafe key. README price: $0.042 per million input tokens, output free. The 12-task bench cost $0.45 of Jev, with a median of $0.004 per task.
Links
FAQ
- Is Quicksilver an official TypeSafe or Anthropic skill?
- README FAQ: no. It is an independent open-source project, and the README says it is not affiliated with Anthropic or TypeSafe AI. The install path is npx github:UditAkhourii/quicksilver, and the key is your own TypeSafe key.
- Does the 86% figure mean every task got cheaper and faster?
- The README states 86% fewer Claude tokens, median 82%, a match on 8 of 12 tasks, up to 20x faster, and a median of $0.004 of Jev. The same table shows the BGL log falling from 54% F1 to 23%, and the 3,000-line needle task moving from 56s to 64s. Huge scans save tokens. They do not always save wall-clock time.
- How does the bench count a Claude token?
- bench/README.md: Claude-alone totals subtract a 68.5k fixed floor from a control agent. Quicksilver is charged the full SKILL.md on every situation, plus each command, the full stdout, and 40 wrapper tokens. The count is characters divided by 4. Agents on the Claude-alone side used Sonnet.
- What should stay on Claude or on grep?
- README: writing, editing, multi-step reasoning, and anything grep answers exactly. Let Claude check the ? rows on look-alike code and on house-style labels. S12 is the numeric negative control, and bench/README.md says plain awk wins there.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- Jev & TypeSafe
A primer for builders who need to know what Jev is, who it is for, and when TypeSafe docs should take over from this directory.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.