Introducing jevfetch: an open source web reader for coding agents
An open source tool that began as a side project at a2sys. Give it a URL and a question, and it hands your agent the relevant passages verbatim.
Meet jevfetch, an open source tool that began as a side project at a2sys.
AI agents frequently need to read the web, whether to check technical documentation or find a specific condition buried in a long page. Reading an entire page for just a few relevant sentences takes time and adds processing overhead. Getting a summary from another AI can mean revisiting the source to verify exact wording or important conditions.
jevfetch improves this process. Give it a URL and a question, and it selects relevant passages and delivers them verbatim to the agent. Think of it as highlighting the relevant sections of a long manual before handing it over. It also preserves the heading hierarchy so the agent can understand each passage in context.
Why not WebFetch?
This offers clear advantages over Claude Code's WebFetch for targeted lookups. WebFetch uses a smaller AI model to read a page and formulate a response; jevfetch selects the original passages that support the answer. This is particularly useful when looking for exact numbers, conditions, or feature descriptions in lengthy documentation. It avoids rewriting the selected text and lets the agent that produces the final answer work directly with the source.
Under the hood, jevfetch uses Jev, an AI model from TypeSafe designed for fast judgments such as "Does this passage answer the question?" or "Which option fits best?" Jev selects the relevant passages, while agents such as Claude Code or Codex produce the final answer. Each handles the part it does best.
What we measured
Our internal benchmark on October 23 showed a substantial speed improvement. Across 30 documentation questions with URLs provided, total agent task time was approximately 47% shorter than WebFetch in both Korean and English. Of 116 checks for required facts per language, WebFetch passed all 116 in both languages, while jevfetch passed 115 in Korean and 116 in English.
This was a limited evaluation on specific documents and questions, but the time savings with comparable coverage of required facts were encouraging. Full conditions are in the benchmark notes.
Try it
jevfetch is best suited to finding specific information in long pages, rather than translating or summarizing entire documents. It is available for use with Claude Code, Codex, and other agents, and requires a TypeSafe API key.
npx skills add a2sys-ai/jevfetch --skill jevfetch -a claude-code
This started as a small side project, but it reflects a broader belief: making AI work better requires both smarter models and better tools for delivering the information they need. We will keep improving jevfetch based on practical use and feedback. Give it a try and let us know what you think!
Code and setup guide: github.com/a2sys-ai/jevfetch