Ground any LLM in the live web.

Q: Does Grounder only help small local models?

No. Every model is frozen at its training cutoff. In a first benchmark, grounder improved a local 7B from 3/8 to 7/8 correct and a frontier cloud model from 0/5 to 3/5 on current-fact questions. The token-capped pack additionally helps models with small context windows.

Q: Does Grounder store query content?

No. Grounder processes query content to fetch results but does not store it for telemetry, analytics, training, or profiling. Only billing and abuse counters are kept, and neither contains a word of what you searched.

Every model is frozen at its training cutoff. Grounder gives any LLM three MCP tools - live search, the real page, and a token-capped cited evidence pack - so it answers from current sources instead of memory. Your model writes the answer; Grounder stores no query content.

Get a free key

1,500 pages a month free, no credit card.

The tools

A small surface your model can route without misfiring. Full reference in the tools & API docs.

web_search - live Google results: titles, snippets, ranked source URLs. Your model's URL-discovery step - search, then fetch the pages worth reading.
fetch - a page as clean markdown, the actual current page, fetched live. Pass a query to get back only the relevant passages, capped to your token budget.
deep_search - one question, one cited evidence pack. It searches, reads several pages, ranks the passages across all of them, and returns the few that fit your max_tokens - each with its source URL.

Does grounding help?

A first benchmark on current-fact questions - the kind a frozen model gets wrong. The same questions, with and without a grounder evidence pack.

Model	Alone	+ Grounder
Qwen2.5-7B (local, 8k window)	3 / 8	7 / 8
DeepSeek-chat (cloud, large window)	0 / 5	3 / 5

Unaided, the 7B said "Tavily was acquired by ServiceNow" (it was Nebius) and overflowed its 8k window on raw pages; the cloud model got every current fact wrong. A big context window doesn't fix a frozen one.

Directional first run - 8 questions, single seed, ground truth pinned from primary sources. Full method, every arm, and the losses we report rather than bury.

Pricing

One flat monthly price, and every page is one page - a search, a fetch, or each page deep_search reads. You only pay for pages we deliver.

Per month	Free	Starter	Pro
Price	$0	$9	$19
Pages	1,500	5,000	15,000
Protected pages	-	✓	✓

Full pricing, every limit, and how pages work →

Add it to your MCP client

One block, no SDK. Works with Ollama, LM Studio, Continue.dev, Aider, Claude Desktop, and Cursor.

{
  "mcpServers": {
    "Grounder": {
      "command": "uvx",
      "args": ["grounder-mcp"],
      "env": { "GROUNDER_API_KEY": "your_key_here" }
    }
  }
}

Or get a key by email - it's sent back instantly:

Guides

How-tos for wiring Grounder - or a DIY pipeline - into your stack. All guides →

FAQ

Does Grounder only help small local models?

No. Every model is frozen at its training cutoff, so grounding helps all of them. In a first benchmark a local 7B went from 3/8 to 7/8 correct, and a frontier cloud model from 0/5 to 3/5. The token-capped pack additionally helps small context windows, where dumping raw pages overflows and fails.

Does Grounder store my searches?

No. It processes query content to fetch results but stores none of it for telemetry, analytics, training, or profiling. Only billing and abuse counters are kept, and neither holds a word of what you searched.

Is Grounder a Tavily alternative?

Yes, if your priority is cited evidence over an opaque synthesis, a token-capped pack that fits your window, the actual live page instead of a cached index, and one flat monthly price instead of a per-request meter where a single request can quietly cost many times another.