Merged into anymd (2026-09-25). anymd reads PDFs, Office files, EPUB, web pages, images and video into clean Markdown for AI agents:
npx -y @sylphx/anymd. This repository is archived.
Lookout
Web answers with source-level proof.
Search and fetch citeable excerpts. No API key. Hits name the adapter. Fetches keep a span, warnings, and gaps. Multi-step research is a separate tool.
npm @sylphx/lookout · bin lookout · MCP io.github.SylphxAI/lookout
The problem
A search snippet is a title and a URL. The agent then paraphrases a page it never opened. When the body is missing, a confident summary is worse than a gap.
The difference
| A snippet says | Lookout returns |
|---|---|
| “Release notes” and a URL | title, description, excerpt span, and any table rows it could read |
| nothing when the fetch fails | a warning or an error: blocked address, HTTP status, truncation, empty extract |
Fetch reads the HTTP response. It does not run JavaScript, and it does not write a model summary.
Search, then open the page
npx -y @sylphx/lookout search "Model Context Protocol"
npx -y @sylphx/lookout fetch https://example.com
npx -y @sylphx/lookout extract https://example.com
search queries DuckDuckGo HTML, Wikipedia OpenSearch, the npm registry, and Hacker News Algolia, then ranks the hits. It does not fetch those pages and it does not run web_research.
Install
npx -y @sylphx/lookout
That starts a stdio MCP server. No API key.
| Your client | Setup |
|---|---|
| Any agent / CLI | npx -y @sylphx/lookout |
| Claude Code | claude mcp add lookout -- npx -y @sylphx/lookout |
| Claude Desktop / Cursor / VS Code / Codex | "command": "npx", "args": ["-y", "@sylphx/lookout"] |
{
"mcpServers": {
"lookout": { "command": "npx", "args": ["-y", "@sylphx/lookout"] }
}
}
Tools
| Tool | When to call it |
|---|---|
web_search | Ranked hits from the four public adapters. No API key. |
web_fetch | One URL. SSRF checks, redirect cap, size cap, body-prefix span. |
web_extract | Title, metadata, tables, and spans from HTML or a URL. |
web_cache | Query or manage the on-disk cache. Not a web search. |
web_crawl | Bounded same-origin crawl. Not the search default. |
web_research | Named multi-step tool: search, then fetch and extract the top pages (default 3, max 6). |
web_diff | Added and removed words between two texts or two URLs. |
There is no profile switch. The tool you name is the work that runs. Cache is on for search and fetch unless you pass useCache: false. A cache hit replays an earlier response from disk. The first search still needs the network.
What it will not pretend
- Private, loopback, and link-local addresses are rejected.
- A failed adapter, a non-200 response, a truncated body, or an empty extract is reported. It is not turned into a clean answer.
web_fetchdoes not honorrobots.txtunless you setrespectRobots: true.web_crawlhonors it unless you turn that off.- Lookout does not render a browser and does not synthesize an answer.
Companion MCP tools
| Product | Job |
|---|---|
| Citra | PDF answers with page-level proof |
| Iris | Image facts and pixel evidence |
| Cue | Video timelines and timestamp evidence |
| Spine | Repository architecture and impact |
| Locus | Exact code-chunk retrieval |
Each product is independent. Install only the tools your agent needs.
Documentation
| Website | sylphxai.github.io/lookout |
| Quickstart | Install and first call |
| Compare | Firecrawl, Tavily, and what Lookout actually does |
Development
bun install
bun test
bun run doctor
bun run docs:build
License
MIT