deterministic research · mcp server

Rival/Search

A deterministic research server for AI agents — nine tools, five engines, nine social platforms, six academic sources, zero API keys.

Get started → Add to Cursor View on GitHub 95 stars
Powering: academic research
Terminal
$ claude mcp add RivalSearchMCP --url https://RivalSearchMCP.fastmcp.app/mcp
FastMCP 3 Python 3.10+ MIT License No in-server LLM rivalsearchmcp.com
9
// tools
Each annotated with read/write/idempotent hints so clients skip needless confirmation prompts.
28sources
// data hubs
5 web · 9 social · 5 news · 9 academic — concurrent and deduplicated.
0
// api keys
No tokens, no subscriptions, no per-call billing. Paste the URL and go.
0
// in-server llm
No model runs in the server. Output stays deterministic and auditable.
see it work

Watch an agent run the whole research stack.

One prompt. The agent picks the tools, fans out across 28 sources, and hands back scored, structured results — no keys, no black box.

▸Picks the right tool per step// web · social · content_operations
▸Returns markdown + JSON// chain one result into the next
▸Deterministic + auditable// no in-server LLM to hallucinate
agent · RivalSearchMCP
the premise

Agents need to read the internet. They shouldn't need ten servers to do it.

01Single-source MCP servers force agents to juggle brave-search, tavily, arxiv, reddit, github — separate auth, separate quotas, separate failure modes.
02"AI-powered research" servers smuggle a hidden LLM into the tool layer, so results hallucinate, drift, and resist auditing.
03Free tier rarely means free. API keys, subscriptions, and rate caps push research-heavy agents into the four-digits-a-month bracket fast.
// before
$ connect brave-search
$ connect tavily-mcp → enter api key
$ connect reddit-mcp → oauth flow
$ connect arxiv-mcp
$ connect github-mcp → enter token

// after
$ connect rivalsearchmcp
✓ 9 tools registered. ready.
capability surface

Six categories. One connection.

// web
5
DuckDuckGoBingYahooMojeekWikipedia
// social
9
RedditHacker NewsStack OverflowDev.toMediumProduct HuntBlueskyLobste.rsLemmy
// news
5
Google NewsBing NewsThe GuardianGDELTDDG News
// academic
9
OpenAlexCrossRefarXivPubMedEurope PMCKaggleHuggingFaceDataverseZenodo
// code & sites
2
GitHub searchmap_websiteresearch / docs / map
// documents
4
PDFWordPlain textImages (OCR)

28 live data sources across web · social · news · academic — plus code search and document analysis, every one keyless.

tool surface

Nine tools. Three categories. Every result auditable.

Each returns markdown + structured JSON. See the full reference →

01web_searchConcurrent multi-engine search. Scrapling-backed TLS fingerprinting bypasses Cloudflare. Per-engine failures don't block others.search
02social_search9 platforms in parallel — Reddit, HN, Stack Overflow, Dev.to, Medium, Product Hunt, Bluesky, Lobste.rs, Lemmy. No auth.search
03news_aggregation5 news sources merged and time-windowed (day · week · month · anytime). Deduped by URL across engines.search
04github_searchRepository search with built-in rate limiting (60/hr unauthenticated). Optional README inclusion for deeper context.search
05map_websiteSite traversal in research / docs / map mode. Per-page quality scores plus aggregate confidence.search
06content_operationsOne tool, six ops: retrieve · stream · analyze · extract · score · find_conflicts. Returns markdown + structured JSON.analysis
07research_topicTwo modes: topic (search + rank) and entity (8-source fan-out profile of any named entity, in parallel).analysis
08document_analysisPDF, Word, plain text, and images. EasyOCR for image extraction — lazy-loaded, no setup. 50 MB cap.analysis
09scientific_research5 paper providers and 4 dataset hubs. Returns relevance-ranked, citation-scored academic results.workflow
the design principle

The server runs no LLM. Every result is deterministic, every score is reproducible, every output is auditable. The caller's model does the synthesis — because a consistent machine can't hallucinate the way a synthesizing one can.

// auditable
ToolResult ships markdown for humans and structuredContent JSON for agent chaining — no regex parsing.
// scored
Every multi-result tool returns a per-item quality score (0–100) and an aggregate confidence signal.
// reconcilable
find_conflicts surfaces numeric, date, and polarity disagreements as a first-class signal instead of averaging them away.
why builders choose it

The reasons it gets installed — over and over.

Pulled from how it's described across the MCP ecosystem — LobeHub, mcpservers.org, plugable.io, and the Claude & Cursor marketplaces.

// 01

Zero keys, zero setup

"Just add the URL and use all tools immediately." No tokens, no OAuth, no signup — the most-repeated line in every registry listing.

— Awesome MCP Servers
// 02

No black box

No LLM runs inside the server, so nothing hallucinates in the tool layer. Every result is deterministic, reproducible, and auditable.

— core design principle
// 03

One server replaces many

Web, social, news, academic, code, and documents behind a single connection — instead of wiring up five siloed servers with five quotas.

— 28 sources, 1 endpoint
// 04

Free & open, no rate caps

"A free, open-source server with no rate limits or authentication barriers." MIT licensed — use the hosted endpoint or self-host.

— plugable.io
// 05

Built for agent chaining

Every tool returns markdown for humans and structured JSON for machines — pipe one tool's output into the next, no regex parsing.

— structured ToolResult
// 06

Indexed everywhere

Listed across LobeHub, Glama, mcp.so, DXT, and the Claude & Cursor marketplaces — discoverable wherever agent builders look for tools.

— 7+ registries
drop in, anywhere

One URL. Every MCP client. Now a plugin, too.

Quick options below, or follow the full installation instructions → for Cursor, Claude, VS Code, Codex, and self-hosting.

Add RivalSearchMCP to Cursor
CDClaude Desktopremote server✓ ready
❯Cursorone-click install✓ ready
{}VS Code.vscode/mcp.json✓ ready
$_Claude Code+ plugin marketplace✓ ready
OXOpenAI Codexplugin✓ ready
⌂Self-hostuv sync · fastmcp run✓ ready
~/.cursor/mcp.json · .vscode/mcp.json
{
  "mcpServers": {
    "RivalSearchMCP": {
      "url": "https://RivalSearchMCP.fastmcp.app/mcp"
    }
  }
}
claude code · plugin marketplace
$ claude plugin marketplace add damionrashford/RivalSearchMCP
$ claude plugin install rival-search-mcp@rivalsearchmcp
✓ 9 tools registered. ready.
hosted server · one line
$ claude mcp add RivalSearchMCP --url https://RivalSearchMCP.fastmcp.app/mcp
how to use it

Connected? Just ask your agent.

No new syntax to learn. Prompt naturally — RivalSearchMCP picks the right tool and returns clean, cited results you can chain.

›"Use RivalSearchMCP to search recent news about open-source AI agents."
›"Use RivalSearchMCP to compare FastAPI and Django across web, GitHub, and academic sources."
›"Use RivalSearchMCP to retrieve this URL, score the source, and flag conflicting claims."
›"Use RivalSearchMCP to find scientific papers on retrieval-augmented generation."
how we differ

Broad and deterministic. Nobody else sits here.

▸28 sources, one server// rivals wrap a single paid API
▸No in-server LLM// no synthesis = no hallucinated tool output
▸Structured markdown + JSON// chainable; not prose to re-parse
▸Zero keys · free · OSS// every competitor here is paid
RivalSearchMCP28 sources · free · deterministic
brave-search-mcpsingle source · paid key
Exa MCPpaid API · in-server ranking
Tavily MCPpaid · LLM synthesis
Perplexity MCPpaid · black-box answers
built in the open

Real traction. No paid promotion.

// github stars
95
Organic growth — picked up by agent-tooling readers and the major MCP directories on "no keys, no in-server LLM."
// forks
16
Developers cloning to extend adapters and self-host — the clearest signal a build is genuinely reusable.
// sessions
5,000+
Sessions tracked across the site and repo — sustained interest, not a launch-day spike.
// repo clones
650/ 14 days
Builders pulling the repo to self-host and extend — a fresh two-week window.
// repo views
480+ / 14 days
Unique and total GitHub repo visitors over the last two weeks.
// listed in
7+ registries
LobeHub, Glama, mcp.so, DXT and more — plus a SpiderShield "safe" security verdict.
LobeHubGlamaAgentHotspotEliteAI.toolsmcpmarketmcp.soDXT.soSpiderRatingSpiderShield ✓ safe
use it · feature it

Building with RivalSearchMCP? Show it off.

Rival Search MCP is a free, open-source project. If it powers your project, link back and feature the logo — it helps other builders find it.

// contribution guideline

If you use RivalSearchMCP in a public project, please:

  • Link to the GitHub repo and rivalsearchmcp.com
  • Feature the logo or drop the badge below in your README
  • Star the repo so other builders discover it
README.md · badge markdown
[![Powered by RivalSearchMCP](https://rivalsearchmcp.com/assets/rival-search-logo.png)](https://rivalsearchmcp.com)
questions

Everything you'd ask before installing.

What is Rival Search MCP?

An open-source Model Context Protocol (MCP) server — a search MCP that gives AI agents real-time web research across 9 deterministic tools and 28 data sources (web, social, news, academic, code, documents), with zero API keys and no LLM inside the server.

Is it free and open source?

Yes — Rival Search MCP is a free, open-source project under the MIT License. Use the hosted server at no cost, or clone the repo and self-host. No paid tiers, no rate caps.

Does it require API keys?

No. Zero API keys, tokens, or authentication for all 9 tools. Add one URL to your MCP client and you're running.

How do I install it?

Add https://RivalSearchMCP.fastmcp.app/mcp to your MCP client, or run claude mcp add RivalSearchMCP --url …. It also ships as a plugin for Claude Code and OpenAI Codex. See the install guide.

Can I self-host it?

Yes. Clone the repo, run uv sync, then fastmcp run server.py (STDIO) or add --transport http. A one-command fastmcp install also wires it into Claude Desktop, Cursor, and Claude Code.

How is it different from other search MCP servers (Tavily, Exa, Brave)?

Those wrap a single paid API behind a key and often run an LLM in the tool layer. Rival Search MCP spans 28 sources across six categories, is free and keyless, and stays deterministic — no in-server synthesis to hallucinate. Full comparison →

// questions & bugs Got a question, or hit a bug? Open an issue — it's the fastest way to reach the maintainer and it shapes the public roadmap.
Open an issue → Browse issues

Plug it in. Ship a research agent today.

Open source, hosted, and ready. Add the URL, install the plugin, file an issue, send a PR — every contribution lands in a public roadmap.

Connect in 30 seconds Add to Cursor Contribute on GitHub

Like it? Star the repo and follow the creator —

Star Follow @damionrashford Connect on LinkedIn