rozetyp/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.
A thin stdio client for the hosted service at https://grounder.dev - no browser, no scraper,
nothing heavy runs locally. It runs with **no signup** on a shared demo key; add your own free key
(1,500 pages/month, no card) at grounder.dev for real use.
Install
Configure
Claude Desktop, Cursor, LM Studio, Continue.dev, or any MCP client. The key is **optional** - omit
it to try Grounder on the shared demo key, or add your own free key from grounder.dev:
The four tools
| Tool | What it does |
|---|---|
| web_search | Google organic results plus people-also-ask, related searches, and the knowledge graph. The top snippet often already holds the answer. |
| 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, ranked into a token-capped, cited evidence pack. Optional grounded answer. |
| research | Investigates an open question with no ready-made answer: it plans, reads primary sources, notices what is missing, goes back for it, and reasons to a grounded conclusion. |
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 pages can be 20,000+ tokens (we measured 22,759
for one query) - enough to make a small model return nothing. You get the relevant passages, not
whole pages.
- **The live page, not a cached copy.** fetch reads the actual current page; any caching is short,
timestamped, and force-refreshable.
- **Flat monthly price, billed in pages.** You only pay for pages actually delivered. No per-call metering.