pi-terse-web

pi-terse-web

Lean web fetch + web search for the Pi coding agent. Two tools, each one network call:

• webfetch — URL → clean markdown (HTML pages) or raw text, byte- and char-capped.
• websearch — query → Exa /search results as {title, url, snippet}.

The chat model synthesizes from what these return. Neither tool curates, summarizes, opens a browser, or runs a fallback cascade.

Why

pi-web-access is powerful but heavy: seven search providers, a headless curator, a YouTube/PDF/video/GitHub extractor chain, and a DNS-resolving SSRF guard that breaks under fake-IP proxies (Clash/Mihomo TUN). pi-terse-web is the lean opposite for people whose web access is mostly docs, articles, and search results:

• No SSRF DNS gymnastics. guardUrl only checks the URL scheme and validity — it does not resolve DNS and keeps no host/IP denylist, so the fake-IP-proxy class of failures can't happen by construction. This is deliberate for a local single-user agent: the process already owns the network, and an incomplete denylist is worse than none. Network-policy isolation is the sandbox/runtime's job, not this tool's.
• One fetch, one timeout. No Jina → Parallel → Gemini cascade that takes 30s+ and bills multiple API calls.
• Proxy is free under TUN. Clash/Mihomo TUN intercepts at the network layer, so no proxy configuration is needed — the fake-IP SSRF bug class can't happen because there's no DNS guard to begin with. For HTTP_PROXY/HTTPS_PROXY setups, run Node 24+ with NODE_USE_ENV_PROXY=1 (Node's global fetch then follows the env vars). Either way, zero proxy code in the tool.
• Bounded output. A byte-capped stream read prevents huge pages exhausting memory; output is char-capped with a truncation marker, so a fetch can't overflow the context window.

Install

pi install npm:pi-terse-web

Or try it for one run:

pi -e npm:pi-terse-web

Configure

websearch needs an Exa API key:

export EXA_API_KEY=...

webfetch needs no configuration.

Coexisting with pi-web-access

The two can coexist — pi-terse-web uses distinct tool names (webfetch, websearch) vs pi-web-access (fetch_content, web_search), so there is no registration conflict. But the model will see overlapping capabilities; if you want a single lean set, disable pi-web-access's tools or uninstall it.

License

MIT.