@walterra/pi-web-tools

Web research extension for pi coding agent providing:

• web_search: paid Kagi Search API
• web_fetch: Jina Reader Markdown extraction

Compatibility

• Tested with pi 0.84.3 (@earendil-works/pi-coding-agent)

Configuration

Set credentials before starting Pi:

export KAGI_API_KEY='...'

# Optional; defaults to hosted Jina Reader
export JINA_READER_URL='https://r.jina.ai'

# Optional for authenticated/higher-limit hosted Jina usage
export JINA_API_KEY='...'

For a self-hosted Reader, set JINA_READER_URL to its private endpoint. The current web_fetch URL policy only accepts source URLs on ports 80 and 443; this does not restrict the Reader endpoint itself.

Never commit API keys or place them in tool arguments.

Installation

pi install npm:@walterra/pi-web-tools

For local development from this repository:

pi -e ./packages/pi-web-tools/extensions/web-tools/index.ts

Environment changes require restarting Pi because /reload does not change the parent process environment.

Behavior

web_search

• Returns 5 results by default, maximum 10.
• Makes one Kagi API request per uncached invocation.
• Uses a 15-minute in-memory cache.
• Never paginates automatically.
• Set noCache: true only when freshness requires another paid request.

web_fetch

• Uses https://r.jina.ai by default.
• Returns frontmatter plus Markdown.
• Uses a one-hour in-memory cache.
• Supports Reader auto, curl, and browser engines.
• Rejects credentials, non-HTTP protocols, unusual source ports, and source hostnames resolving to private/reserved addresses.
• Marks retrieved text as untrusted external content.
• Truncates output to Pi's tool-output limits.

Verification

Check that the extension loads:

pi --no-extensions -e ./extensions/web-tools/index.ts --list-models >/dev/null