jevfetch skill
Skill instead of WebFetch reads a page and returns only the blocks that answer the question, verbatim with heading paths.
- 55 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Page text is sent to the external Jev service for block judging.
- A platform key is required and sits in an environment variable.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/jevfetchDetects 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.
- Page text is sent to the external Jev service for block judging.
- A platform key is required and sits in an environment variable.
This is third-party code. Review the repository files before installing.
What it does
jevfetch reads web pages for coding agents. It splits a page into blocks, asks TypeSafe's judgment model Jev whether each block answers the question, and returns the matching blocks plus the first one exactly as they appear. The agent gets quotes rather than a paraphrase and can verify a fact without rereading the page. The skill installs through `npx skills add`, and the tool itself through `uv tool install jevfetch`. There is a command line and a Python API. The authors report a 30-question documentation benchmark where agent runs in Claude Code took 47 percent less time than with `WebFetch` at similar cost.
Who it is for. Developers whose agents often hunt for an exact fact on long documentation pages, spending time and context on it.
Good fit when
- An agent checks a specific parameter in the docs and must quote it without paraphrase.
- The page is long and putting it all into context is expensive.
- You want to switch `WebFetch` off and read every page with one tool.
- Questions about documentation come in Korean and English.
Not a fit when
- You need a summary or translation of the whole page, since blocks are picked for a question.
- Pages cannot be sent to an outside service.
- You have no TypeSafe key and do not want to get one.
Example request
Read the documentation page at this address and quote the paragraph about request limits verbatim, with its heading path.Limitations
It needs Python 3.12 or newer, `uv`, Node.js for `npx` and a TypeSafe key in `TYPESAFE_API_KEY`. Page blocks go to the Jev service for judging, so private documents should not be read this way. The authors' measurements apply to specific Claude Code versions and October 2026 dates. On the Korean set jevfetch passed 115 of 116 checks and `WebFetch` passed all 116. Codex blocks command network access by default, so you must allow it in `~/.codex/config.toml`.
How to disable. Remove the skill from `.claude/skills/` and the `Bash(jevfetch:*)` rule from settings, then uninstall the command with `uv tool uninstall jevfetch`. If you disabled `WebFetch`, lift the deny rule in `.claude/settings.json`.
FAQ
How does it differ from WebFetch?
`WebFetch` returns a small model's paraphrase, jevfetch returns quoted blocks with heading paths. That is easier to verify for an exact fact on a long page.
Is a key required?
Yes, a TypeSafe key in `TYPESAFE_API_KEY`. It calls the Jev model that decides which blocks answer the question.
How do I install without Node.js?
Download the Claude Code skill file by hand into `~/.claude/skills/` or the project's `.claude/skills/`, and install the command with `uv tool install jevfetch`.
Can I call it from code?
Yes, the package has a command line and a Python API, described in the README sections.
What if Codex cannot reach the network?
Codex CLI and the IDE extension block network access for commands by default. Allow it in `~/.codex/config.toml`, otherwise installation and every call fail.
Discussion
No questions or reviews yet.
Related
SkillsA skill that turns a link into clean Markdown via Scrapling with a Jina Reader fallback, incl. WeChat, Juejin and CSDN articles
webctl searches the web for an agent, scores results with Jev and returns only useful chunks to save context.
A web research skill: SearXNG searches, Camofox reads pages and CloakBrowser opens sites that check their visitors
A skill for web search, X search and page fetching for models without web access: no signup or key, the answer is JSON with source links
An MCP server and CLI that crawls sites: a small model picks links and passages so the agent reads only the answer.
A skill that reads a link and hands the agent the full text as Markdown, picking a separate route for WeChat pages, Feishu docs, PDFs and arXiv papers
README in short
The README explains the block principle, gives benchmarks against `WebFetch`, steps for Claude Code and Codex, terminal usage and the Python API.