grounder-mcp

# Grounder MCP

Live web grounding for local and cloud LLMs, as four MCP tools. Every model is frozen at its
training cutoff; Grounder gives yours the current web - ranked results, real page content, and a
cited evidence pack sized to your context window.

This is a thin stdio client for the hosted service at **[grounder.dev](https://grounder.dev)** - no
browser, no scraper, nothing heavy runs locally. Get a key (free, no card) at grounder.dev.

## Install

```bash
uvx grounder-mcp          # or: pip install grounder-mcp
```

## Configure

Claude Desktop, Cursor, LM Studio, Continue.dev, or any MCP client:

```json
{
  "mcpServers": {
    "grounder": {
      "command": "uvx",
      "args": ["grounder-mcp"],
      "env": { "GROUNDER_API_KEY": "gnd_live_your_key" }
    }
  }
}
```

## The four tools

| Tool | What it does |
|---|---|
| **`web_search`** | Google organic results plus the surfaces around them - people-also-ask, related searches, knowledge graph. The top snippet often already holds the answer, so the model can skip a fetch. |
| **`fetch`** | One page as clean markdown, capped to your token budget, plus the final URL after redirects. |
| **`deep_search`** | One search, read across the pages it surfaces, chunked and ranked into a token-capped, cited **evidence pack**. Optional grounded answer, written only from what it read. |
| **`research`** | An agentic search-and-read loop for hard, multi-part questions: it plans, reads primary sources, notices what is still missing, goes back for it, and answers from the pages it actually read. |

## Why use it

- **It fits a small context window.** Results come back token-capped, so they slot into an 8-32k
  local model instead of overflowing it. A few raw web pages can be 20,000+ tokens - we measured
  22,759 for one query - which is enough to make a small model return nothing at all. `deep_search`
  and `research` hand back the relevant passages, not whole pages.
- **The live page, not a cached copy.** `fetch` reads the actual current page on request. Any caching
  is short, timestamped, and force-refreshable - it never turns into an opaque stale index.
- **Flat monthly price, billed in pages.** One page is one search, one fetch, or one page a
  `deep_search` reads, and you only pay for pages actually delivered - an empty search or an
  unreadable page is free. No per-call metering.
- **No query content stored.** Ever.
- **Nothing heavy to run.** The client just forwards tool calls to the hosted API, so it installs in
  seconds with no browser or scraping stack on your machine.

## Pricing

Free: 1,500 pages/month, email only, no card. Starter $9/mo and Pro $19/mo add higher page budgets,
protected-page access, and more `research` runs. Full table at
[grounder.dev/pricing](https://grounder.dev/pricing).

## Links

- Site and docs: [grounder.dev](https://grounder.dev) · [docs](https://grounder.dev/docs) · [guides](https://grounder.dev/guides)
- MCP registry: `io.github.rozetyp/grounder`
- PyPI: [grounder-mcp](https://pypi.org/project/grounder-mcp/)

<!-- mcp-name: io.github.rozetyp/grounder -->