Nimrod runs on your computer and gives Claude, or any MCP app, Google search with dates and filters, page and PDF reading that hands your AI the content itself — not a summary of it — and multi-source research. One command to install, a browser sign-in, one predictable bill.
Your AI reads the page itself: its main content, PDFs included, in pieces
sized for a context window. Search shows what's fresh, and research recovers when a
source won't open. Already installed? Run nimrod update.
Pages come back as their main content — the article, docs or post — with menus, ads, cookie banners and sidebars left out. Tables keep their columns, formulas come through once, links stay clickable.
Papers, datasheets and reports are read like any other page, on your computer, up to 25 MB.
A long page arrives 20,000 characters at a time, with a map of its sections, so your AI jumps straight to the part it needs instead of flooding its context.
One call reads up to ten pages in parallel and returns each page's text — in about a third of a second in our tests.
Every result with its site and date, plus the questions people ask next. Filter from the last hour to the last year, by country and language, or search news, images and video.
Sources follow Google's ranking, two per site at most. When a page is blocked or empty, the next result takes its place. Stack Overflow and Rust crates are read through their official APIs.
Some pages only
exist once a browser has run their JavaScript. Turn on
scrape_webpage with nimrod scrape on
(or the "Scrape pages" switch in Claude Desktop) and your AI can have such a page
rendered remotely. It costs more than a search: 1 credit for most pages, up to
5 for pages that need a full browser. Reading further into a page you've scraped is
free, and a scrape that fails costs nothing.
The same five research questions, answered by the same AI model, once with Nimrod and once with typical built-in web tools. Both got all five right.
Tested 2 October 2026 with Claude Sonnet: a software release, a news event, two vendors' prices, figures from a research paper and a git how-to. The built-in fetch tool summarizes each page with a second model, which once reported a 2026 date as 2024; Nimrod returns the page's own text. That text costs more tokens than a summary — we think accurate is worth it. Sites that block automated readers, like AP and Reddit, blocked both.
Because raw search is where agent runs go to die. Nimrod exists to keep your model reasoning instead of wading.
Google's own ranking, with each result's site and date, clean URLs, and the questions people ask next. Pages come back as their main content — the article, docs or paper, PDFs included — with menus, ads and cookie banners left out, so Claude reasons over evidence instead of wading through page furniture.
Pages arrive as their main content in markdown — no nav bars, cookie banners, or 90 KB of markup soup. Long pages come in 20,000-character parts with a map of their sections, so one page can't swamp the conversation and your AI reads only the part it needs.
research_topic runs search → dedupe → read →
synthesis in one call: Google's ranking, two sources per site at most, and the next
result standing in for any page that won't open. What normally takes a dozen
improvised tool hops becomes a single structured step.
Site, phrase, date (last hour to last year), language and country filters. One credit = one search, reading pages always free — costs you can predict before the run starts. Works in Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, Gemini CLI, and LM Studio.
Searches go through Nimrod; web pages and PDFs are fetched and read by Nimrod on your computer, never by our servers. Everything speaks the Model Context Protocol — current spec (2026-07-28) with full backward compatibility.
Google results with site, date and snippet, plus related questions and searches. Site, phrase, date, language, country and news/image/video filters.
1 credit
A page or PDF to its main content as markdown, text or HTML, with its dates. Long pages come in parts with a section map. Read on your computer.
free
Up to ten pages or PDFs read in parallel on your computer, each page's text returned. Pages that fail say why.
free
Searches, then reads the top 3, 5 or 8 sources on your computer — swapping in the next result when one won't open — and hands Claude a structured synthesis prompt.
1 credit / search
Renders a page remotely in a full browser, for pages built with JavaScript that
can't be read on your computer. Costs more than a search, so your AI only reaches for
it when reading fails. Turn on with nimrod scrape on.
One credit is one Google search. Reading pages is free on every plan:
it runs on your computer. Optional page scraping, for pages that need a full browser,
costs 1 to 5 credits per page.
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.
One command. Setup signs you in through your browser — no key to copy — and adds Nimrod to whichever of Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, Gemini CLI, and LM Studio you have.
curl -fsSL https://nimrod.orchis.ai/install.sh | sh
irm https://nimrod.orchis.ai/install.ps1 | iex
Builds aren't code-signed yet.
The install commands avoid security prompts. If you download a file by hand, Windows
may show SmartScreen (More info → Run anyway), and on a Mac the
.mcpb may be blocked by Gatekeeper — use the install
command there instead. Binaries, checksums, and older versions:
GitHub releases.
Add a custom connector (Settings →
Connectors) with https://nimrod.orchis.ai/mcp and sign
in with your email. It offers google_search and a
search-only research_topic. Web page extraction runs
locally only, so it isn't available on web or mobile. Install Nimrod on your
computer to read pages.
In Claude Code, Nimrod also installs skills, an agent, and hooks that shape how Claude researches — same account, 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-read-verify method — not just a prompt.
Multi-source research runs in its own isolated context: the agent searches, reads, 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. Other apps get the same advice when they connect.
| Nimrod on your computer | Connector on claude.ai | |
|---|---|---|
| Works in | Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Codex, Gemini CLI, LM Studio | claude.ai on the web and mobile |
| Google search with dates and filters | ✓ | ✓ same credits |
| Reading web pages | ✓ free, on your computer | — runs locally only |
| Reading PDFs | ✓ free, on your computer | — runs locally only |
| Optional page scraping | ✓ 1–5 credits per page, off by default | — |
| Research synthesis | ✓ reads the top pages | from search snippets |
| Skills, agent, and hooks | ✓ in Claude Code | — |
Manage it from a terminal:
nimrod status, nimrod update,
nimrod scrape on|off, and
nimrod uninstall, which removes it from every app.
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.
So the pages you read are fetched by software on your machine, at your
direction, and never pass through our servers — which is also faster and more
private. Our servers only run your searches and keep track of credits. The one
exception is optional scraping (off unless you run
nimrod scrape on): it has a page that needs a full
browser rendered remotely, at 1 to 5 credits per page. On
claude.ai web and mobile, which can't run local tools, the connector offers
search only. Comparison above.
One credit is one Google search executed on your behalf. Reading web pages
is free because it runs on your computer — 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. Optional scraping
(scrape_webpage, off by default) costs 1 credit for
most pages and up to 5 for pages that need a full browser; reading further into a
page you've scraped is free.
No. nimrod setup (or your first search) signs you
in through your browser with your email. API keys still work for scripts and CI:
get one in the portal. We store only a hash of each key, so it's shown once — if
you lose it, rotate it in the portal.
For a page your AI needs, yes; for bulk harvesting, no. The optional
scrape_webpage tool renders one page at a time when
it can't be read on your computer, at 1 to 5 credits a page. 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. If you need
firehose-scale data collection, we're honestly not the right tool.
Yes, since 2.1: papers, datasheets and reports up to 25 MB, read on your computer like any other page and free. Long ones come in parts with a section map. Scanned PDFs without a text layer can't be read.
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. The pages you read are fetched by Nimrod on your computer and never reach our servers — except pages you choose to scrape, which pass through without being stored. Details in the privacy policy.
The current revision (2026-07-28) with automatic backward compatibility for older clients, both in the app on your computer and in the claude.ai connector.