MCP Omnisearch is a Model Context Protocol (MCP) server that puts several web search, answer, and extraction services behind four tools, so an agent calls one server instead of one per provider. Each provider is enabled by its own API key; providers without a key are skipped and the rest keep working.
The four tools split the work by task rather than by vendor. websearch runs a web search through Tavily, Brave, Kagi, or Exa, with controls such as result limit, search depth, topic, time range, safe search, and raw content that apply when the selected provider supports them. aisearch returns sourced answers from Kagi FastGPT, Exa Answer, Linkup, or Tavily Research, where Tavily Research first returns a task ID that you pass back as researchid to collect the report. githubsearch searches GitHub code, repositories, or users. web_extract extracts, crawls, scrapes, summarizes, or finds similar content through Tavily, Kagi, Firecrawl, or Exa.
The server is TypeScript and runs from the published npm package or a local build. FIRECRAWLBASEURL points the Firecrawl provider at a self-hosted instance, and OMNISEARCHLARGERESULT_MODE chooses whether large results are returned inline or written to a file. Accompanying documentation covers provider selection by task and capability, a search-operator support matrix, large-result behavior, deployment, and troubleshooting.
Features
- Four consolidated tools: websearch, aisearch, githubsearch, and webextract cover search, sourced answers, code search, and extraction
- Provider choice per call: name the provider in the request rather than running one server per service
- Key-gated providers: Tavily, Kagi, Brave, GitHub, Exa, Linkup, and Firecrawl each activate from their own environment variable
- Partial configuration: one key is enough; unconfigured providers are skipped rather than failing the server
- Search controls: limit, search depth, topic, time range, safe search, raw content, and automatic parameters where supported
- Extraction modes: extract, crawl, scrape, summarize, and find-similar, with markdown output and per-source chunking
- Self-hosted Firecrawl: FIRECRAWLBASEURL redirects extraction to your own deployment
- Large-result handling: results can be returned inline or written to a file for the client to read
