A local web research stack for coding agents with no keys and no spin

What it does

wigolo runs as a local MCP server that plugs into coding agents like Claude Code, Cursor, or VS Code. It exposes a durable surface for web tasks—search, fetch, crawl, structured extraction, caching, similarity search, and autonomous research loops—so the agent can ask the web questions without ever calling a cloud API. Everything lives under ~/.wigolo/, including on-device embedding models and a cross-encoder reranker.

The interesting bit

The project treats transparency as a feature, not a bug: it surfaces failed fetches, stale cache hits, weak results, and engine degradation directly in the response rather than papering over them. Its search tool fans out queries across 18 direct engine adapters in parallel, then fuses and reranks results locally with explainable scores—something the README claims serial host tool-loops cannot replicate.

Key highlights

• Core search, fetch, crawl, and cache tools require no API keys and no cloud egress; optional answer synthesis can use a free Gemini tier or a local Ollama instance.
• Fetching auto-escalates through HTTP, TLS impersonation, and a headless browser only when anti-bot measures or SPA shells demand it.
• Research and agent tools decompose questions into sub-queries, gather sources, and synthesize cited reports with step logs and time budgets.
• A built-in benchmark against paid tools (Tavily, Exa, and a built-in web search) claims parity on a Postgres replication query, with the added detail of byte-offset citations and self-flagged junk results.
• All data, models, and configuration stay local by default; the project is AGPL-3.0 licensed.

Caveats

• The project is in public beta, and the README notes it requires Node ≥ 20 and roughly 1.5 GB of disk space for local models.
• Research, agent, and answer-synthesis features need an external LLM provider configured unless you route them through a local Ollama server.
• The benchmark is a single live query reported by the project itself, not an independent or reproducible test suite.

Verdict

Developers who want their AI agents to read the web without managing API budgets or shipping data to third parties should look here. If you need a managed SaaS with SLAs and a support line, this is explicitly not that.