Nimrod is a hosted MCP server for Claude: quality-scored Google search, clean webpage extraction, and deep multi-source research. One URL to add, one key, one predictable bill — no scraping stack to babysit.
Because raw search is where agent runs go to die. Nimrod exists to keep your model reasoning instead of wading.
Every result is scored for quality, rated for authority, and classified by source type — code, docs, news, forum, web — before your model spends a single token on it. Claude reasons over ranked evidence instead of guessing which of ten blue links is content-farm filler.
Pages arrive as boilerplate-stripped markdown — no nav bars, cookie banners, or 90 KB of markup soup. A fraction of the tokens of raw HTML means cheaper turns, faster answers, and more context left for actual thinking.
research_topic runs search → rank → dedupe →
extract → synthesis in one metered call. What normally takes a dozen improvised
tool hops — each one a chance for the agent to wander — becomes a single
structured step.
Site, date, and language filters. One credit = one search, extraction always free — costs you can predict before the run starts. Works in every MCP client: Claude Code, Claude Desktop, claude.ai.
Everything speaks the Model Context Protocol — current spec (2026-07-28) with full backward compatibility.
Organic Google results with quality scores, authority ratings, and source-type classification. Site, language, date, and vertical filters.
1 credit
Any URL to clean markdown, text, or HTML — boilerplate stripped, summary included.
free
Up to five URLs fetched and summarized concurrently for fast source scanning.
free
Searches, ranks, deduplicates, and extracts sources, then hands Claude a structured synthesis prompt. Three depth levels.
1 credit / search
One credit is one Google search. Extraction is free on every plan.
Checkout by Polar, merchant of record.
Every plan is limited to 30 requests per minute. Nimrod is a research tool,
not a scraping stack — the cap is deliberate and applies across the board.
Three ways in — all live today.
claude mcp add nimrod \ --transport http \ https://nimrod.orchis.ai/mcp \ --header "Authorization: Bearer nmk_…"
https://nimrod.orchis.ai/mcpnimrod-desktop setup, thenclaude mcp add nimrod -- path/to/nimrod-desktopThe hosted connector gives Claude four good tools. Nimrod Desktop turns them into a research practice: skills, an agent, and hooks that shape how Claude researches — running on your machine, on the same key and the same credits.
One-command workflows Claude already knows how to run:
/deep-research, /fact-check,
/find-docs, /news-monitor,
/competitive-analysis, /cite-sources.
Each encodes a full search-extract-verify method — not just a prompt.
Multi-source research runs in its own isolated context: the agent searches, extracts, cross-checks, and returns a cited brief — instead of flooding your conversation with raw page dumps.
Before every metered call, Claude automatically gets query-crafting guidance:
exact phrases, site and date filters, when one research_topic
beats five searches. Your credits go further without you thinking about it.
Skills, the agent, and
hooks are Claude Code features (they install into your ~/.claude
setup). The four research tools themselves work in every client, including
Claude Desktop via nimrod.mcpb.
| Hosted connector | Nimrod Desktop | |
|---|---|---|
| Works on claude.ai, mobile, anywhere | ✓ | Claude Code & Claude Desktop |
| Quality-scored Google search | ✓ | ✓ same credits |
| Webpage extraction | ✓ free, runs on our server | ✓ free, runs on your machine — the pages you read never transit our servers |
| Research synthesis | ✓ | ✓ assembled locally |
| Six research skills | — | ✓ |
| Deep-research agent | — | ✓ |
| Search-coach hooks | — | ✓ |
| Extra cost | none | none — one download, same key |
Download:
nimrod.mcpb for Claude Desktop (Windows) ·
Windows ·
macOS ·
Linux for Claude Code.
Skills, agent, and hooks install themselves on first run and remove cleanly with
nimrod-desktop uninstall.
Nimrod is the research half of a live-data toolkit. Run it beside your other MCP tools and your agent stops relying on a stale training cutoff.
Design hardware with an agent that checks reality: Konnect drives your schematics and PCBs while Nimrod pulls current parts availability, datasheets, errata, and app notes from the live web — no guessing from year-old training data.
Coding agents verify library versions and breaking changes before refactors. Analyst agents cite today's filings, not last year's. Any MCP-capable tool that acts on the world pairs with a research tool that can see it.
The connector is the fastest way in and all you can use on claude.ai and mobile. The desktop install adds everything that can't live behind an endpoint: six research skills, a deep-research agent, and search-coach hooks install into your Claude setup, and webpage extraction runs on your machine — free, faster, and private, since the pages you read never pass through our servers. Same key, same credits, no extra cost. Full comparison above.
One credit is one Google search executed on your behalf. Webpage extraction
is free because it doesn't cost us search inventory — we only meter what costs
money to serve. research_topic charges one credit
per underlying search it runs: 1 for the main topic plus 1 per focus area you
add. Its depth setting changes how many sources are read, which is
free — depth never costs extra credits.
We store a cryptographic hash of your key, never the key itself — so a database breach can't leak working credentials. The flip side: we can't show it again. If you lose it, rotate in the portal: the old key dies, a new one is issued and shown once.
No — and the limits make that structural: every plan, free through the largest pack, is capped at 30 requests per minute. Nimrod is built for research agents that read a handful of good sources per question, not for bulk harvesting. If you need firehose-scale data collection, we're honestly not the right tool.
Free-trial credits are valid for 30 days from signup (one-time, no renewal). Monthly plan credits reset each calendar month. Purchased pack credits are valid for 6 months from purchase and are drawn only after any monthly allowance — so nothing you paid for is wasted while a subscription is active.
Polar is the merchant of record: they process the card, handle VAT/sales tax globally, and manage refunds. We never see your payment details.
We keep per-call metering records (tool, credits, latency) to run the service and show you your own usage. We don't sell usage data, and extraction requests fetch pages directly — no third-party analytics in the path. Details in the privacy policy.
The current revision (2026-07-28, stateless) with automatic backward compatibility for older clients — Claude Code, Claude Desktop, and claude.ai all connect with no configuration.