A local tool that lets coding agents read and research the web without cloud costs or key setup
It reduces the costs and data leakage concerns that accumulate every time an agent calls external APIs for web search and crawling. All processing happens on your computer, and results include precise source citations, allowing the agent to build answers based on reliable evidence. This is a tool that processes the AI agent's web information collection and analysis locally.
How the agent interacts with the web
This tool bundles all processes of an AI agent interacting with the web into a single interface. It provides ten tools: search, page fetching, site crawling, data extraction, caching, similar page finding, research, and autonomous collection loops. Because the browser engine and ML models are installed and run on your device, no API keys are required and data is not sent to cloud servers. No additional costs are incurred regardless of how much the agent explores the web, and all data remains within the ~/.wigolo/ folder to ensure privacy.
The form of results received by the agent
Search results are not just a list of links but are composed of evidence that the agent can use immediately. Each result includes an excerpt with the byte offset of the phrase in the original text precisely specified, a citation ID, and score breakdown information. Weak results or failed engines are automatically filtered or clearly marked, so the agent always knows what evidence it is standing on. Cached data is reused immediately, providing responses without latency for repeated queries.
Installation and agent connection
Node.js 20 or higher and approximately 1.5GB of disk space are required. The npx wigolo init command sets up the local engine, and adding the --agents option completes the connection with major coding agents like Claude Code, Cursor, and Codex in one go. It operates as an MCP server, integrating naturally into existing agent environments, and can also connect to other frameworks or self-hosted agents via REST API or SDK. You can check the system status at any time with the npx wigolo doctor command.
Differences from paid services
This tool is designed as an alternative to paid web search APIs, but it is not a simple replacement; it is an agent-specific web layer. It queries 18 search engines in parallel and applies rank fusion and ML reranking to improve precision. Pages with bot blocking are automatically handled by switching to a headless browser, and failures are reported with clear causes. In benchmark tests, it derives the same core answers as paid services while providing excerpts fixed at the byte level and explainable score breakdowns.
Things to check before use
It is currently in public beta and is being verified for stability with over 7,600 test switches. An LLM provider key is required to generate synthesized answers in the research and agent tools, but core functions like search and crawling work without a key. If using a data center IP, access may be restricted due to IP reputation checks by some sites, so proxy configuration should be considered in such environments. For commercial use, a commercial license is required under the AGPL-3.0 license.
