analog-mcp

A Model Context Protocol server for Analog — Analog lets any AI understand websites as code — usable from Claude Desktop, Cursor, and anything else that speaks MCP. The perception layer for LLMs: webpages in, structured records out. Deterministic, zero-shot, honest. The extraction pipeline contains no LLMs.

Pages are fetched on your machine by default. The SDK uses a local headless browser, which visits as an unauthenticated user by default. For structured extraction, the SDK sends the URL and rendered page content to Analog. Results return to the SDK and are saved locally unless disabled. Authenticated content is sent only when you explicitly provide it through html= or a custom fetcher. mode="local" is the Markdown-only path that sends nothing.

This server is Analog for MCP clients. Agents with shell access can use the analog command-line app directly — the same extraction and saved-result workflow as shell commands, no configuration required.

The headline tool is analog(url). Point it at a webpage; get back a bounded receipt with section shapes, representative records, completeness disclosures, and a handle for the saved result. Its JSON text fallback never exceeds 16,384 characters; structured content carries the same receipt.

Tools

• analog(url, pages=N) — extract a page; returns a bounded orientation receipt plus a handle. Pass pages=N (max 50) to follow the site's own next-page links and merge up to N pages into one result (each page costs one extraction; the sweep is disclosed on the response).
• assess(url) — check which tool fits a page for your task: the page's measured markdown size plus task-keyed guidance (working with the data → analog; one quick question → read it directly; a platform data feed spotted → worth knowing), so you route a page without spending an extraction. No sign-in required.
• analog_history(limit=N) — list saved results, newest first, as cheap metadata (handle, url, when saved, size) to recover a handle you no longer have in context.
• analog_open(handle) — return a previous result in full, no re-fetch.
• analog_export(handle, fmt) — re-emit it as json / csv / yaml / md.
• analog_section(handle, index) — pull one section's records.
• analog_collection(handle, index) — query compatible physical sections as one identity-deduplicated records view while retaining their provenance.
• analog_find(handle, pattern, field=) — search a result's records across sections; pass field= to scope the search to one field.
• analog_describe(handle) — a token-cheap preview of a saved result (section shapes plus sample records) for orienting without pulling the full JSON.
• analog_feedback(...) — report a page that extracted poorly (reason labels are enumerated in the tool's schema) or send a free-form feature request.

Every analog result is saved locally, so the analog_* tools re-slice it without re-fetching or re-rendering the page.

JSON-shaped tools expose navigable MCP structured content and an equivalent text fallback. Format exports, previews, and feedback confirmations remain text-only.

Install

The server reads the credential stored by the analog CLI (it ships with analog-sdk), so install both tools:

uv tool install analog-sdk
uv tool install analog-mcp

Prefer pip? pip install analog-mcp in one environment puts both executables on PATH. uvx analog-mcp runs the server one-off, but signing in still needs the CLI installed.

Sign in

Connect a new account:

analog signup

Or connect an existing account:

analog login

This opens your browser, signs you in, and stores the key in the CLI's local credential store — analog-mcp reads it from there, so no key ever lands in an MCP config file.

Analog renders pages with a built-in headless browser that downloads on first use (~150 MB, one-time). To fetch it ahead of time so the first analog tool call doesn't pause:

analog browser install

Configure your MCP client

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "analog": {
      "command": "analog-mcp"
    }
  }
}

Restart Claude Desktop. The analog tool will appear in the tool list. No env block is needed — analog-mcp reads the credential you stored during sign-in.

Cursor / other MCP clients

Same idea — the server command is analog-mcp and it speaks stdio. Refer to your client's MCP configuration docs for the exact file location.

Usage

Once configured, just ask the agent things like:

Extract the speakers from https://example-conference.com/2026

The agent calls analog(url=…) under the hood. The tool returns a bounded receipt with representative records and a saved-result handle. The agent can answer from that receipt or use the handle to request the complete result, one section, a search, or another format without fetching the page again.

Full documentation: https://getanalog.io/docs

What this version ships

analog-mcp exposes the extraction tool analog(url), a local routing check assess(url), and the analog_* re-slice tools, over the stdio transport.

License

MIT — see LICENSE.

Links

• Homepage: https://getanalog.io
• Command-line app and Python SDK: analog-sdk on PyPI