Quicksilver skill
A Claude Code skill that hands bulk sorting of files, logs and tickets to the fast Jev model to cut Claude token use
- 142 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- File and log content goes to the third-party TypeSafe API, so secrets must not end up in requests.
- The Jev key is stored in `~/.quicksilver/config.json` with user-only permissions.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/quicksilverDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- File and log content goes to the third-party TypeSafe API, so secrets must not end up in requests.
- The Jev key is stored in `~/.quicksilver/config.json` with user-only permissions.
This is third-party code. Review the repository files before installing.
What it does
Quicksilver installs a skill and a `qs` utility into Claude Code. They take over work where a lot of material is read to make a small decision: which of 187 files touch authorization, which log lines hold real failures, which of 200 tickets ask for a refund. Those questions go to the Jev model from TypeSafe, and Claude receives a short list with a confidence mark. The `qs` utility offers `filter` (pick files or lines by a yes or no condition), `classify` (sort items into given labels), `rank` (order by relevance), `find` (locate a spot in a huge file) and `ask` (put a question with state from a file). Borderline results carry a question mark so Claude checks them itself. After every run a receipt shows items scanned, Jev cost and Claude tokens not read. The skill makes Claude phrase one narrow judgment instead of asking it to read everything and report the important parts. The author publishes a reproducible measurement over 12 tasks in the `bench/` folder.
Who it is for. Developers with large repositories and logs in Claude Code whose context fills up with reading done only to pick out the relevant pieces.
Good fit when
- Among hundreds of files you must find those that handle sessions or authorization.
- A log of thousands of lines has to be reduced to real failures.
- Incoming tickets or reviews are sorted into given categories.
- The project is huge and the question of where request retry lives is settled by semantic ranking.
- A short list of suspicious files is needed before a full security review.
Not a fit when
- The task is writing or editing text and code or multi-step reasoning, which Jev cannot do.
- An exact `grep` gives the answer.
- The content must not be sent to a third party.
Example request
Find files in the src folder that handle user sessions and give a short list with confidence for each.Limitations
A Jev key from console.typesafe.ai and Node 18 or newer are required. File content goes to the `api.typesafe.ai` API, so secret data must not be sent there. The utility never sends `.env`, private keys, certificates or credentials files, respects `.gitignore` and skips binaries and files over 2 MB. In the author's measurement quality matched Claude on 8 of 12 tasks. On look-alike code and subjective labels such as commit types the result is worse, so question-mark items need checking. This is an independent project with no link to Anthropic or TypeSafe.
How to disable. Delete the `~/.claude/skills/quicksilver` folder and restart Claude Code. The key is removed by `npx github:UditAkhourii/quicksilver setup --remove` (the `~/.quicksilver/config.json` file).
FAQ
How do I install Quicksilver in Claude Code?
Run `npx github:UditAkhourii/quicksilver`: the installer copies the skill into `~/.claude/skills/quicksilver` and asks for a Jev key once. Plugin install also works: `/plugin marketplace add UditAkhourii/quicksilver`, then `/plugin install quicksilver@quicksilver`.
How much does Jev cost?
Input tokens cost $0.042 per million and output is free. The whole 12-task measurement cost the author $0.45.
Does Quicksilver replace Claude?
No. Jev cannot write, edit or reason. The skill only keeps selection-driven reading out of Claude's context.
Which tasks should the skill avoid?
Writing, editing, multi-step reasoning and anything `grep` answers exactly. On huge scans the gain is in tokens, not time.
Where does the data go?
File content is sent to the `api.typesafe.ai` API. TypeSafe states that Jev is not trained on customer data. The utility never sends `.env` files, keys or certificates.
Discussion
No questions or reviews yet.
Related
SkillsA skill that makes Claude a manager: it writes tickets, hands code to cheaper agents, verifies and asks another vendor to review.
A plugin that cuts token use in long sessions: prunes stale tool output, picks a subagent model and narrows the skill list.
Thirteen skills hand an agent's small decisions to the cheap Jev model: model, skill, memory passages and next step.
A Claude Code plugin with hooks and 26 skills: measures what the agent knows, blocks edits until code is read, keeps memory
A hundred ready loops for Claude Code, Cursor and Codex: the agent acts, runs a check and fixes the cause until the goal is met.
A Claude Code skill that hands implementation-heavy work to Codex CLI while Claude keeps planning, review and acceptance.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Quicksilver (this card) | no | 142 | |
| Caveman | no | 119,585 | shortens the agent's text answers; Quicksilver hands another model the selection work Claude would otherwise read |
| RTK | no | 86,384 | compresses terminal command output; Quicksilver keeps whole-file and whole-log reading out of the context |
README in short
The measurement covers 12 tasks: a supercomputer log, Banking77 intent routing, UCI SMS spam filtering, SST-2 sentiment, the Hono codebase and commit history, and lodash. The author compares a Claude Code subagent with the same work done through Quicksilver against hidden ground truth.