omnisearch

Free open-source software. Use entirely at your own risk.
No warranties. No support contract. $0 liability cap. You alone own API keys, credit burn, supply-chain risk, and anything that lands in a merged PR.
Full terms: DISCLAIMER.md (required reading). Apache-2.0 — LICENSE.

See named accounts and OmniRoute gateways for multiple accounts, gateway-only credentials, supported authentication and health.

Search MCP — parallel search across many engines so your AI finds what a single search usually misses.

• Parallel by default. Every engine you configure is queried at once, not one at a time.
• Merges and dedupes. One ranked list. Duplicate URLs and titles collapse. Each hit records which engines found it.
• Web, news, code, and social. Brave and GitHub join automatically when keyed. News, X, Reddit, YouTube, and more when you add their keys.
• Research mode. Search first, then read the top pages so answers can cite real sources.
• Drop-in for Claude Desktop and Cursor. Point the client at the omnisearch binary and start searching.

Apache-2.0. Copyright 2026 Ravi Tharuma. This is gratuitous software (typically $0 to the authors) — not a paid product. See DISCLAIMER.md §0 and §5.

Community

Doc	Purpose
DISCLAIMER.md	Assumption of risk, $0 liability, supply chain, spend, indemnity
CONTRIBUTING.md	Dev setup, DCO, contributor warranties, PR rules
docs/reporting-issues.md	What to include (and redact) in issues
SUPPORT.md	Where to get help
SECURITY.md	Private vulnerability reporting
CODE_OF_CONDUCT.md	Community standards
GOVERNANCE.md	Maintainership and releases
CHANGELOG.md	User-visible changes
NOTICE	Apache NOTICE

Get started

Clone the repo, add the keys you have, run the binary:

git clone https://github.com/RaviTharuma/omnisearch.git
cd omnisearch
cp .env.example .env
omnisearch

omnisearch talks to Claude Desktop, Cursor, and any other MCP client over stdio. Use omnisearch http if you want a local HTTP endpoint instead.

If the binary is not on your PATH yet, install it from this checkout (cargo install --path .) or drop a release build on your PATH. Toolchain notes are under Development.

Release binaries

Tagged releases provide native binaries for four targets:

Platform	Architecture	Target
Linux	x86_64	x86_64-unknown-linux-gnu
Linux	ARM64	aarch64-unknown-linux-gnu
macOS	Intel	x86_64-apple-darwin
macOS	Apple Silicon	aarch64-apple-darwin

Download omnisearch-v0.2.2-<target>.tar.gz and SHA256SUMS from the GitHub Release. Verify the archive with sha256sum --check --ignore-missing SHA256SUMS on Linux, or shasum -a 256 --check --ignore-missing SHA256SUMS on macOS, then extract it and put omnisearch on your PATH. Archives include the binary, license, and README. Linux builds use Ubuntu 24.04 and require a compatible glibc runtime; these are not static musl binaries. Windows release binaries are not currently provided.

The release workflow accepts v* tags only when the tag exactly matches the package version in Cargo.toml and Cargo.lock. All four builds run on native runners with --locked and pass the local-only MCP stdio smoke before upload. No release is published until every target succeeds.

Cursor

.cursor/mcp.json:

{
  "mcpServers": {
    "omnisearch": {
      "command": "omnisearch",
      "args": ["stdio"],
      "env": {
        "BRAVE_API_KEY": "...",
        "GITHUB_TOKEN": "...",
        "TAVILY_API_KEY": "tvly-...",
        "EXA_API_KEY": "..."
      }
    }
  }
}

Claude Desktop

Same command and env, no args:

{
  "omnisearch": {
    "command": "omnisearch"
  }
}

Any other MCP client can use command: omnisearch with args: ["stdio"].

HTTP (optional)

AUTH_TOKENS=replace-me OMNISEARCH_HTTP_BIND=127.0.0.1:48731 omnisearch http

http://127.0.0.1:48731/mcp with Authorization: Bearer replace-me. OMNISEARCH_HTTP_RPM rate-limits per token.

Tools

Tool	Purpose
search	Query every configured engine in parallel and return one merged list. Default mode=all. auto picks engines from the query and recent health. ladder tries free engines first.
ai_search	Answer-oriented subset (Tavily / Kagi / You.com / Exa / Perplexity when configured).
research	Search, then extract top URLs under a time budget.
extract / web_extract	Pull page text from extract vendors, then a safe direct fetch.
brave_search tavily_search exa_search linkup_search kagi_search	Single-engine web search.
github_search	Repos (kind=repo), code (kind=code), users (kind=users).
x_search reddit_search discord_search youtube_search instagram_search facebook_search	Official social APIs.
firecrawl_scrape firecrawl_crawl firecrawl_map	Site scrape / crawl / map.
get_provider_info	Configured, cost, requires_key, notes. No secrets.
search_health	Cooldown, latency, recent errors.
quality_report	Domain diversity and coverage.
provider_bench	Latency / error bench.

search arguments:

{
  "query": "Swiss AI regulation 2026",
  "providers": ["brave", "github", "tavily", "exa"],
  "limit": 20,
  "unlimited": false,
  "mode": "all",
  "search_type": "news",
  "freshness": "week",
  "country": "CH",
  "language": "de",
  "max_providers": 8,
  "timeout_seconds": 20,
  "budget_usd": 0.05,
  "no_cache": false,
  "cache_partial": false,
  "ground_top": 3,
  "evidence_min": 8,
  "quality_report": true
}

Each hit: title, url, snippet, provider, score?, confidence?, published_at?, sources[], snippet_grounded.

Run meta: selected, successful, failed, timed_out, skipped, cost_usd, provider_used, stop_reason. Partial runs are not cached unless cache_partial is set.

limit is a per-engine hint. Omit it or set unlimited: true to keep paging until engines are exhausted or 10,000 unique results. Budgets cap spend, how many engines run, and time — not a tiny result ceiling.

Providers

Unconfigured engines are skipped. Wikipedia, Semantic Scholar, Bluesky, and Reddit public JSON work without keys. Brave (BRAVE_API_KEY) and GitHub (GITHUB_TOKEN / GITHUB_API_KEY) join the default run when set.

Provider	Env	Search	Extract	Notes
Brave	BRAVE_API_KEY	yes	—	Web + news. On by default when set.
GitHub	GITHUB_TOKEN or GITHUB_API_KEY	yes	—	Repos, code, users. On by default when set.
Tavily	TAVILY_API_KEY	yes	yes	News topic + time_range
Exa	EXA_API_KEY	yes	yes	Neural / keyword + /contents
Firecrawl	FIRECRAWL_API_KEY	yes	yes	scrape / crawl / map
Linkup	LINKUP_API_KEY	yes	—	outputType=searchResults
Kagi	KAGI_API_KEY	yes	—	Authorization: Bot …
Perplexity	PERPLEXITY_API_KEY	yes	—	POST /search
You.com	YOU_API_KEY or YDC_API_KEY	yes	yes	ydc-index.io/v1/search
Parallel.ai	PARALLEL_API_KEY	yes	yes	/v1/search, /v1/extract
Querit	QUERIT_API_KEY	yes	yes	Multilingual filters
TinyFish	TINYFISH_API_KEY	yes	—	Source-only
Keenable	KEENABLE_API_KEY or KEENABLE_PUBLIC=true	yes	keyed	Public tier is keyless
X	X_BEARER_TOKEN or XAI_API_KEY	yes	—	API v2 recent search, else xAI x_search; multi-account via OMNISEARCH_ACCOUNTS
Reddit	none, or OAuth pair	yes	—	search.json; OAuth when set; public {} or OAuth named accounts
Discord	DISCORD_BOT_TOKEN + DISCORD_GUILD_IDS	yes	—	Guild message search only; optional DISCORD_AUTH=bearer
YouTube	YOUTUBE_API_KEY or GOOGLE_API_KEY	yes	—	Data API v3
Instagram	Graph token + business account	hashtag	—	Official hashtag API only
Facebook	FACEBOOK_ACCESS_TOKEN	pages	—	Official /pages/search only
Wikipedia	none	yes	—	MediaWiki
Scholar	none	yes	—	Semantic Scholar
Mastodon	MASTODON_INSTANCE	yes	—	/api/v2/search
Bluesky	none	yes	—	app.bsky.feed.searchPosts
MCP backends	OMNISEARCH_MCP_BACKENDS	yes	—	`name

OMNISEARCH_MCP_BACKENDS=official attaches remotes for every configured key (and named OMNISEARCH_ACCOUNTS api_key pools): Tavily, Exa, Firecrawl, Linkup, Kagi, Perplexity. Dual/triple env keys become tavily / tavily_2 / … Named accounts become tavily-work style backend names and replace legacy keys for that provider. Brave remains native must-have fan-out, not an official remote.

Social API limits: Instagram is hashtag-only (Meta caps unique hashtags 30 / 7 days). Facebook is Pages Search only. X prefers X_BEARER_TOKEN on https://api.x.com/2/tweets/search/recent. Discord has no public global search — configure guild IDs; Bot tokens are often rejected on Discord’s message search route (use a user bearer or expect the provider to fail/skip).

How it works

• Every configured engine runs in parallel. mode=ladder runs free engines first and can stop once enough evidence is in (stop_reason=evidence).
• Results merge with reciprocal rank fusion: score += 1 / (k + rank) (OMNISEARCH_RRF_K, default 60). Confidence blends that score with recency and domain trust.
• Tracking parameters are stripped. sources[] lists every engine that returned the URL.
• Dual-key failover: NAME_2 / NAME_3 after 429 / 5xx / timeout.
• Health cooldown: 429 / timeouts cool an engine; mode=auto skips it.
• ground_top fetches the top N URLs (private and loopback hosts blocked) and reframes the snippet.
• mode=auto ranks engines by recent success and query type.
• OMNISEARCH_AUTO_ALLOW_USD skips expensive engines unless you list them in providers[].
• In-process cache with a TTL. Partial runs are not cached unless asked. no_cache: true bypasses.
• Known shorteners are dropped; hits per domain are capped, then overflow is appended.
• Freshness: day / week / month / year. Defaults OMNISEARCH_COUNTRY=CH, OMNISEARCH_LANGUAGE=de.
• Extract and direct-fetch block loopback, link-local, private IPs, file:, and metadata hosts.
• Payloads over OMNISEARCH_INLINE_MAX_BYTES spill to a temp file.

Development

Needs a Rust 1.88+ toolchain. Binary and crate name: omnisearch. v0.1.0.

cargo install --path .
cargo fmt
cargo clippy --all-targets -- -D warnings
cargo test
cargo build --release
omnisearch bench --query "Swiss AI regulation 2026"