Fixed Deep research and web fetch in Claude Code : full open source

GrapeRoot released Webify MCP, an open-source tool that provides AI coding agents with adaptive web research capabilities at 91% of Deep Research quality for 5% of the cost, integrating with Claude Code, Cursor, VS Code, Windsurf, and Zed.

Adaptive web research for AI coding agents 91% of Deep Research quality · 5% of the cost · Works in every MCP client A skill by GrapeRoot https://graperoot.dev Docs: 中文 /kunal12203/webify-mcp/blob/master/docs/README.zh-CN.md · 日本語 /kunal12203/webify-mcp/blob/master/docs/README.ja.md · 한국어 /kunal12203/webify-mcp/blob/master/docs/README.ko.md · Español /kunal12203/webify-mcp/blob/master/docs/README.es.md · हिन्दी /kunal12203/webify-mcp/blob/master/docs/README.hi.md · Français /kunal12203/webify-mcp/blob/master/docs/README.fr.md · Deutsch /kunal12203/webify-mcp/blob/master/docs/README.de.md · Português /kunal12203/webify-mcp/blob/master/docs/README.pt-BR.md · Русский /kunal12203/webify-mcp/blob/master/docs/README.ru.md pip install webify-mcp claude mcp add webify -- webify-mcp That's it. Works in Claude Code, Cursor, VS Code, Windsurf, and Zed. php flowchart TB Agent AI Agent -- |"web find 'query' "| Webify Agent -- |"web lookup url, 'query' "| Webify Webify -- |"80–300 tokens"| Agent subgraph Webify Webify MCP Server Search Search\nBrave / DDG -- Graph DOM Structural\nGraph Builder Graph -- Retrieve BM25 + BFS\nRetrieval Retrieve -- Synthesize Haiku\nSynthesis end style Webify fill: 1a1a2e,stroke: 16213e,color: fff style Agent fill: 0f3460,stroke: 16213e,color: fff Two tools for web research — both dramatically cheaper than reading full pages: | Tool | When to use | Cost | |---|---|---| web find query | Research questions, anything needing search | ~$0.003/query | web lookup url, query | You know the exact URL | ~$0.0005/query | php flowchart LR A Query -- B Search\nBrave / DDG B -- C1 Page 1 B -- C2 Page 2 B -- C3 Page 3–6 C1 -- D DOM Graph\n+ BM25 C2 -- D C3 -- D D -- E Multi-aspect\nextraction E -- F Haiku\nsynthesis F -- G "Answer\n ~800 tokens " Adapts depth to query complexity. Simple questions hit 3 sources. Multi-dimensional research scales to 6+ with independent sub-aspect retrieval. Call it multiple times with focused sub-queries for deep-research-level coverage. php flowchart LR A URL + Query -- B Fetch page B -- C DOM structural\ngraph C -- D BM25 scoring D -- E BFS traversal E -- F "Relevant subtree\n 80–300 tokens " Scores nodes against your query, returns only the relevant subtree — 80–300 tokens instead of the 3,000–15,000 tokens of full page text WebFetch puts in context. Blind A/B test against Claude's Deep Research — 15 unseen queries, randomized order, Sonnet judge scoring accuracy + completeness + specificity 1–5 each . | Webify | Deep Research | | |---|---|---| Quality | 68/75 · 91% | 73/75 · 97% | Cost/query | ~$0.003 | ~$0.05+ | Latency | 30–90s | 80–280s | Cost efficiency | 18× better | baseline | Webify finds correct information every time. The gap is always completeness — Deep Research reads more. For most queries that difference doesn't matter; for exhaustive research, call web find multiple times. Per-query breakdown | Query | Webify | Deep Research | |---|---|---| | Battery degradation | 13/15 | 15/15 | | OAuth vs OIDC | 13/15 | 15/15 | | Coral reef bleaching | 14/15 | 15/15 | | CRISPR gene editing | 15/15 | 13/15 | | Earthquake & tsunamis | 13/15 | 15/15 | Once installed, the AI automatically uses Webify for web research instead of expensive built-in tools — no configuration needed. The preference policy is embedded in the package itself. What are the tradeoffs between Raft and Paxos consensus? → Claude calls web find — searches, builds graphs, synthesizes answer Look up rate limits in the GitHub API docs → Claude calls web lookup — fetches that page, returns relevant sections only pip install webify-mcp claude mcp add webify -- webify-mcp Add to your MCP config: { "mcpServers": { "webify": { "command": "webify-mcp" } } } Config file locations: Cursor → ~/.cursor/mcp.json Windsurf → ~/.windsurf/settings.json VS Code / Continue → ~/.continue/config.json Zed → ~/.config/zed/settings.json Command: webify-mcp Transport: stdio pip install --upgrade webify-mcp | Env var | Required | Description | |---|---|---| ANTHROPIC API KEY | For web find | Haiku synthesis | BRAVE SEARCH API KEY | Recommended | Reliable search · | WEBIFY CACHE DIR ~/.cache/webify Search: Brave API if key set → DuckDuckGo Lite free fallback, no key needed macOS / Linux — add to ~/.zshrc or ~/.bashrc : export ANTHROPIC API KEY="sk-ant-..." export BRAVE SEARCH API KEY="BSA..." Windows PowerShell : Environment ::SetEnvironmentVariable "ANTHROPIC API KEY", "sk-ant-...", "User" Environment ::SetEnvironmentVariable "BRAVE SEARCH API KEY", "BSA...", "User" In your MCP config applies only to Webify : { "mcpServers": { "webify": { "command": "webify-mcp", "env": { "ANTHROPIC API KEY": "sk-ant-...", "BRAVE SEARCH API KEY": "BSA..." } } } } Get your keys: - Anthropic → https://console.anthropic.com/settings/keys https://console.anthropic.com/settings/keys - Brave Search → https://brave.com/search/api/ https://brave.com/search/api/ Build a graph for a URL python -m webify build https://docs.python.org/3/library/json.html Look up specific info python -m webify lookup https://docs.python.org/3/library/json.html "parse JSON string" python import webify Multi-source search result = webify.web find "how does mTLS work in service meshes" print result "content" synthesized answer print result "sources" {url, title, confidence, tokens} Single-page lookup result = webify.smart lookup "https://docs.python.org/3/library/json.html", "parse JSON" print result "content" relevant sections only ~376 tokens webify-mcp test server Ctrl+C to exit ls ~/.cache/webify/ check cache → Run webify-mcp: command not found pip install webify-mcp Tool not showing up → Restart your editor after adding to config→ Set web find errors ANTHROPIC API KEY → DDG rate-limited; set web find returns no results BRAVE SEARCH API KEY