Pi extension package that adds bounded web search and page extraction tools.

pi-local-websearch uses local system CLIs instead of hosted API keys:

• web_search wraps ddgr.
• web_fetch wraps trafilatura with XML output.

Early MVP.

Implemented now:

• web_search tool using ddgr --json
• web_fetch tool using trafilatura --xml --with-metadata
• default USA search region (us-en)
• model-visible output caps
• full fetched XML and derived text artifacts in /tmp/pi-web-* when fetch output is capped
• URL safety checks for fetches: only http/https, no embedded credentials, blocks localhost/private/link-local/reserved IP targets

Install the runtime CLIs first and make sure they are available on the same PATH used by Pi:

ddgr --version
trafilatura --version

Install from npm with Pi:

pi install npm:pi-local-websearch

Alternatively, install directly from GitHub:

pi install git:github.com/vihu/pi-web

Then restart Pi, or run /reload in an existing session if your Pi version supports extension reloads.

For project-local installation instead of global installation:

pi install -l npm:pi-local-websearch

For local development from a checkout:

git clone https://github.com/vihu/pi-web.git
cd pi-web
npm install
pi -e .

To use a local checkout globally without installing from GitHub, add its absolute path to Pi settings or run:

pi install /absolute/path/to/pi-web

After installation, Pi exposes two tools to the agent.

Search the web with ddgr and return compact ranked results.

Important defaults and limits:

• default results: 8
• max results: 25
• default region: us-en (USA)
• default safe search: enabled
• default timeout: 10000ms
• DuckDuckGo bang queries beginning with ! are rejected

Common parameters:

{
  "query": "Pi coding agent packages",
  "maxResults": 8,
  "site": "github.com",
  "region": "us-en",
  "timeRange": "m",
  "safeSearch": true,
  "timeoutMs": 10000,
}

Fetch and extract readable text from one HTTP(S) URL with trafilatura --xml.

Important defaults and limits:

• default timeout: 20000ms
• default model-visible cap: 12000
• max model-visible cap: 50000
• comments excluded by default
• tables included by default
• if output is capped, full artifacts are written to:
  • /tmp/pi-web-*/document.xml
  • /tmp/pi-web-*/document.txt

Common parameters:

{
  "url": "https://example.com/article",
  "maxChars": 12000,
  "timeoutMs": 20000,
  "includeComments": false,
  "includeTables": true,
  "includeLinks": false,
  "includeImages": false,
  "fast": false,
}

pi-local-websearch is intended for public web access. It does not add a first-use confirmation gate.

web_fetch blocks obvious SSRF-style targets before invoking trafilatura, including localhost, private networks, link-local addresses, reserved documentation ranges, multicast, and IPv6 local ranges. It also resolves hostnames and rejects DNS results that point at blocked IP ranges.

Pi packages run with full system access. Review package source before installing third-party packages.

npm install
npm run typecheck
npm test
pi -e .