knockoutez/wigolo/wigolo

wigolo

Local-first web intelligence MCP server for AI coding agents. Ships ten tools over stdio. All network results land in a local knowledge cache.

Host-LLM synthesis (read me first)

Wigolo has no internal LLM. It returns structured evidence so the calling model (you) writes the final answer. Fold structure into your reply rather than collapsing it away:

• search with format: "highlights" — ML-scored passages + citations. Quote and cite [N].
• research — when MCP sampling is unavailable (common), the output carries a brief with topics, highlights, key_findings. Use it as the scaffold for the report you write.
• find_similar — may return a cold_start string. Pass it to the user; it explains why results came from the web and how to warm the cache.
• extract with mode: "structured" — one call for tables + <dl> definitions + JSON-LD + chart hints + key-value pairs.
• fetch metadata — surfaces og_type, canonical_url, and og_image; use canonical_url to dedupe tracked/canonical URLs.

Quick Setup

Claude Code:

claude mcp add wigolo -- npx wigolo

Cursor / VS Code / any MCP client:

{
  "mcpServers": {
    "wigolo": {
      "command": "npx",
      "args": ["wigolo"]
    }
  }
}

Warmup (recommended, one-time):

npx wigolo warmup          # installs browser engine + bootstraps search engine
npx wigolo warmup --all    # also installs Firefox, WebKit, ML reranker, and embeddings
npx wigolo warmup --force  # wipe search engine state and rebuild

Warmup flags: --force, --all, --reranker, --firefox, --webkit, --embeddings, --no-searxng, --verify.

Tools

fetch

Fetch a single URL and return clean markdown. Use when you already have a specific URL.

Parameters:

• url (string, required)
• render_js: "auto" (default) | "always" | "never"
• use_auth: boolean (default false) — reuses the user's browser session
• max_chars: number
• section: string — return only the content under a heading
• section_index: number (default 0) — which heading match when multiple hit
• screenshot: boolean (default false)
• headers: object
• force_refresh: boolean — bypass cache
• actions: array of {type, selector, text, ms, timeout, direction, amount} — click, type, wait, wait_for, scroll, screenshot. Forces browser rendering when present.

Example:

{ "url": "https://react.dev/reference/react/useState", "section": "Parameters" }

Tip: section is much cheaper than reading the full page. Repeat fetches of the same URL are free from cache unless force_refresh: true.

search

Search the web and return extracted markdown per result. Use when you don't have a URL yet.

Parameters:

• query (string OR string[], required) — array runs variants in parallel and dedupes
• max_results: number (default 5, cap 20)
• include_content: boolean (default true)
• content_max_chars: number (default 30000)
• max_total_chars: number (default 50000)
• time_range: "day" | "week" | "month" | "year"
• include_domains / exclude_domains: string[]
• from_date / to_date: ISO YYYY-MM-DD
• category: "general" | "news" | "code" | "docs" | "papers" | "images"
• language: string
• search_engines: string[] — override engine selection
• format: "full" (default) | "context" (token-budgeted string) | "highlights" (ML-scored passages + citations) | "answer" (synthesized via MCP sampling; falls back to highlights when unsupported) | "stream_answer" (answer + phase progress notifications)
• max_highlights: number (default 10) — cap when format: "highlights"
• force_refresh: boolean

Example:

{ "query": ["react server components patterns", "RSC data fetching", "react server components streaming"], "category": "docs", "include_domains": ["react.dev"], "max_results": 5 }

Tip: keyword queries beat natural-language questions. A 3–5 item query array usually finds more unique sources than one longer query.

crawl

Crawl a site starting from a seed URL.

Parameters:

• url (string, required)
• strategy: "bfs" (default) | "dfs" | "sitemap" | "map" (URL-only discovery, no content)
• max_depth: number (default 2)
• max_pages: number (default 20)
• include_patterns / exclude_patterns: regex string[]
• use_auth: boolean (default false)
• extract_links: boolean (default false) — returns inter-page link graph
• max_total_chars: number (default 100000)

Example:

{ "url": "https://docs.python.org/3/library/", "strategy": "sitemap", "max_pages": 30, "include_patterns": ["^https://docs\\.python\\.org/3/library/asyncio"] }

Tip: strategy: "sitemap" is faster and more complete than BFS on doc sites. strategy: "map" returns URLs only — cheap way to scope before targeted fetches.

cache

Search previously fetched content without hitting the network.

Parameters:

• query: full-text search — supports AND, OR, NOT, "exact phrase"
• url_pattern: glob (e.g. "*react.dev*")
• since: ISO date
• stats: boolean — returns total URLs, size, date range
• clear: boolean — deletes matching entries (requires one of query, url_pattern, since)
• check_changes: boolean — re-fetches matching URLs, reports changed/unchanged with diff summaries

Example:

{ "query": "useState OR useReducer", "url_pattern": "*react.dev*" }

Tip: cache hits are instant and cross-session. Run this before search or fetch when you suspect the content is already on disk.

extract

Structured extraction from URL or raw HTML.

Parameters:

• url OR html (one required; url wins if both provided)
• mode: "metadata" (default) | "selector" | "tables" | "schema" | "structured" (tables + <dl> definitions + JSON-LD + chart hints + microdata/data-attr/grid key-value pairs in one call)
• css_selector: string — required for mode: "selector"
• multiple: boolean (default false) — return all matches, selector mode only
• schema: JSON Schema object with properties — required for mode: "schema"

Example:

{ "url": "https://example.com/product", "mode": "schema", "schema": { "type": "object", "properties": { "price": { "type": "string" }, "name": { "type": "string" }, "sku": { "type": "string" } } } }

Tip: mode: "schema" does heuristic matching over CSS classes, ARIA labels, microdata, and JSON-LD — no LLM call required. mode: "structured" returns every structured pattern on the page (tables, definitions, jsonld, chart_hints, key_value_pairs) in one response — prefer it over chaining multiple extract calls.

find_similar

Find pages related to a URL or a free-text concept.

Parameters:

• url OR concept (one required)
• max_results: number (default 10, cap 50)
• include_domains / exclude_domains: string[]
• include_cache: boolean (default true)
• include_web: boolean (default true)

Example:

{ "url": "https://react.dev/reference/react/useState", "max_results": 8, "include_domains": ["react.dev", "developer.mozilla.org"] }

Tip: uses hybrid 3-way RRF fusion — keyword search + semantic embeddings + live web search. Each result carries match_signals with embedding_rank, fts5_rank, and fused_score. If the cache is empty or embeddings aren't set up, the response includes a cold_start string — pass it to the user to explain why results came from the web.

research

Multi-step research pipeline with decomposition, parallel search, and cited synthesis.

Parameters:

• question (string, required)
• depth: "quick" (~15s, 2 sub-queries, 5–8 sources) | "standard" (~40s, default) | "comprehensive" (~80s, 7 sub-queries, 20–25 sources)
• max_sources: number (cap 50) — overrides depth default
• include_domains / exclude_domains: string[]
• schema: JSON Schema — if present, report is structured to fill these fields
• stream: boolean — emit progress notifications per phase

Example:

{ "question": "How do modern JS bundlers tree-shake ESM vs CJS?", "depth": "standard", "include_domains": ["webpack.js.org", "rollupjs.org", "esbuild.github.io", "vitejs.dev"] }

Tip: research checks cache internally — no need to pre-probe. With MCP sampling, the tool synthesizes the report directly. Without sampling (the common case), the output ships a brief with topics, highlights, and key_findings, plus the raw sources — the host LLM writes the final report from the brief.

agent

Natural-language data gathering. Plans queries and URLs from a prompt, runs them in parallel within budget, optionally applies a schema.

Parameters:

• prompt (string, required)
• urls: string[] — seed URLs to include
• schema: JSON Schema — extract structured fields per page and merge
• max_pages: number (default 10, cap 100)
• max_time_ms: number (default 60000, cap 600000)
• stream: boolean

Example:

{ "prompt": "Compare pricing tiers for Supabase, Firebase, and Clerk", "schema": { "type": "object", "properties": { "provider": { "type": "string" }, "free_tier": { "type": "string" }, "paid_start": { "type": "string" } } }, "max_pages": 12 }

Tip: output includes a steps array showing every action (plan, search, fetch, extract, synthesize) with timings. Use this to debug why an agent run produced a weak result.

diff

Compare two versions of a page. Point it at a live URL and its cached copy, two URLs, or two markdown blobs.

Parameters:

• old (object, required): one of { url, markdown, content_hash }
• new (object, required): one of { url, markdown }
• output: "unified" (default), "hunks", or "summary"
• granularity: "line" (default), "word", or "section"

Example:

{ "old": { "url": "https://docs.example.com/api" }, "new": { "url": "https://docs.example.com/api" }, "output": "hunks", "granularity": "section" }

Tip: same URL on both sides diffs the cached copy against a fresh fetch. Use output: "summary" for counts only.

watch

Monitor a page for changes over time. Lazy execution — checks run when you call check or when an overdue job is picked up during another tool run.

Parameters:

• action (string, required): "create", "list", "check", "pause", "resume", "delete"
• url or urls: single-URL or batch create (mutually exclusive)
• interval_seconds: required for create (min 60)
• selector: create-only CSS selector to scope the diff
• notification: "inline" (default) or an SSRF-guarded webhook URL
• job_id: required for check/pause/resume/delete

Example:

{ "action": "create", "url": "https://nodejs.org/en/blog", "interval_seconds": 21600 }

Tip: webhook destinations are SSRF-guarded — a job cannot be pointed at internal or loopback addresses.

Workflow Patterns

Quick routing:

• Use when search — you need information but don't have a URL.
• Use when fetch — you already have the URL.
• Use when crawl — you need multiple pages from one site.
• Use when cache — you want to check whether something is already on disk.
• Use when extract — you need specific fields, tables, or metadata, not the whole page.
• Use when find_similar — you have a good page/concept and want related content.
• Use when research — a question needs decomposition and multi-source synthesis.
• Use when agent — a natural-language task needs multi-step data gathering.
• Use when diff — you need to see what changed between two versions of a page.
• Use when watch — you want to monitor a page for changes over time.

Cache-first lookup. Before any fetch or search, probe the cache.

cache({ "query": "oauth2 pkce", "url_pattern": "*auth0.com*" })
// empty? fall through to search
search({ "query": "oauth2 pkce flow", "include_domains": ["auth0.com"] })