v0.1.3

Selection and documentation release. jevfetch keeps more of the text that qualifies an answer, scores fewer blocks twice, and the README shows a WebFetch comparison measured on 2026-10-23.

uv tool install jevfetch

• A block right before or after a kept block under the same heading is kept from a stage-2 score of 0.6 (Config.neighbor_threshold), so a condition next to the answer stays with it.
• Note, tip and warning boxes written as <aside> are now read and scored instead of dropped. --no-notes skips them on pages too long to pass through.
• When a table row is kept, the table's header row is added, so the agent can tell what each column means.
• Stage 2 now rescores blocks from 0.6 instead of 0.3 in stage 1, and starts on each window as soon as that window is scored.
• The agent skill tells agents not to add the answer they expect or extra keywords to the question, and lists the spaces jevfetch puts around superscripts and subscripts.
• The README introduces jevfetch as a reader for coding agents such as Claude Code and Codex. Its WebFetch comparison was rerun on 30 questions with Claude Code 2.1.309, and two questions whose sites serve Markdown that WebFetch passes through whole were removed from the set. bench/RESULTS.md adds a comparison on Exa's WebCode Highlights set.
• The README has a logo and a how-it-works animation, and its images now load on PyPI.

v0.1.2

Documentation release. The jevfetch code is the same as v0.1.1; this release refreshes the README shown on PyPI.

uv tool install jevfetch

• The README, the agent skill, and the Korean README install jevfetch from PyPI instead of a git tag.
• The README leads with the comparison against WebFetch on Haiku 5.5, gives the measurement date and Claude Code version, and adds the tool-level result: the reference sentence verbatim in 81 of 81 outputs, against 25 of 81 for one WebFetch call.
• The skill no longer contradicts itself about finding every occurrence of a term, and drops a repeated line.
• SECURITY.md asks for the installed version and states what is in scope. Private vulnerability reporting is on.

v0.1.1

First release on PyPI. The jevfetch code is the same as v0.1.0.

uv tool install jevfetch

• Published from GitHub Actions with Trusted Publishing; each file carries a provenance attestation.
• The release workflow runs the test suite on the tagged commit before building.
• The source distribution leaves out the benchmark data, the benchmark tests, and uv.lock.
• README links are absolute, so they work on the PyPI page.

v0.1.0

First release.

jevfetch fetches a web page and prints only the passages that answer your question, quoted word for word with their section headings. Use it when an agent needs one fact from a long page and reading the whole page would waste context.

Install

uv tool install git+https://github.com/a2sys-ai/jevfetch@v0.1.0

Or add the skill to Claude Code, Codex and other agents. It installs the command the first time it is needed:

npx skills add a2sys-ai/jevfetch --skill jevfetch -a claude-code

You need a TypeSafe API key in TYPESAFE_API_KEY and Python 3.12 or later.

What it does

• Reads HTML, Markdown, plain text, JSON and XML. Links come back as [text](url).
• Renders the page with headless Chrome when the static HTML has almost no text and Chrome is installed.
• Saves PDFs and other binary files to a temp folder and prints the path.
• Caches Jev responses in ~/.cache/jevfetch.
• Ships an agent skill (skills/jevfetch/SKILL.md) and a benchmark suite (bench/).

Against Claude Code's WebFetch on 30 documentation questions in Korean and English, agent runs took 43 to 49% less time at about the same cost with Haiku 5.5 as the reader, and 32 to 34% less time at 43 to 44% lower cost with Haiku 4.5. Setup and caveats are in bench/RESULTS.md.

MIT licensed. The benchmark's page snapshots and quoted sentences are third-party content; see bench/pages/SOURCES.md.