Kestrel Search

Local web search, page extraction, and relevance ranking for AI coding agents.

Kestrel Search

Kestrel Search gives coding assistants a local, keyless web-search capability.

It searches DuckDuckGo, Bing, or Yahoo, retrieves readable page content, and re-ranks results with BM25. It works from the command line and provides structured JSON output for agents and scripts.

Use it when an agent needs current documentation, release information, standards, or other public web content without leaving the local development workflow.

Why Kestrel Search?

• No search API key required
• DuckDuckGo, Bing, and Yahoo support
• Concurrent search and page retrieval
• Main-content extraction
• BM25 relevance ranking
• Clean JSON output for agents
• Agent skill support for Codex, Claude Code, and GitHub Copilot

Quick example

kestrelsearch search "python dataclasses"

For structured results:

kestrelsearch search "python dataclasses" --output json

Next steps

• Install Kestrel Search
• Run your first searches
• Connect it to your coding agent
• Explore search options

Learn how it works

• Search pipeline
• Ranking and result fairness
• JSON output schema
• Limitations

Project benchmark

In a small paired benchmark run on 28 August 2026, Kestrel’s three-provider fanout completed all 24 trials with answer quality close to native Web Search, lower median latency, and fewer median model tokens. The suite contained eight technical research tasks with three trials per task.

Read the full results and limitations. This project benchmark is not a universal performance claim or a DeepSearchQA leaderboard score.