SpecStream's search MCP tops WideSearch at 60.9, beating every other search API
Xageee · reddit · 2026-11-08
SpecStream founder Tarek introduced an MCP server for agent web search, scoring 60.9 on ByteDance Seed's WideSearch with GPT-5.6 Luna — above the best published scores of Parallel (57.6), Exa (55.9), Tavily (55.9) and Perplexity (53.5), and 99% on SimpleQA Verified. Notably, a smaller model on their search beats bigger models on competitors' APIs.
How it works: no proprietary index — each call reads the live web, querying several search engines in parallel, reading result content, opening top pages when sources disagree, then one model pass picks the answer with source URL and read time attached.
Three MCP tools: searchweb (natural-language question, answer plus backing rows, seconds per call), readpages (1–20 URLs per call incl. PDF/Word/Excel, textless pages not billed), and Deep Search (structured rows of every live-web match).
Setup & pricing: one JSON block for Cursor (.cursor/mcp.json), custom connector for Claude/ChatGPT. $5 per 1k searches, $0.80 per 1k pages, $10 free credit without a card.
More from coding & agent
- Twombly launches: a shared canvas for collaborative design with AI agents — floguo · 2026-11-08
- GraphRAG vs RAPTOR vs HippoRAG: how to pick a RAG architecture for AI agents — MaryamMiradi · 2026-11-08
- Dev builds anti-hallucination research agent that verifies 108 citations word-for-word — ApprehensiveMark1308 · 2026-11-08
- Interview with Microsoft CoreAI EVP Jay Parikh on Copilot Fusion and AI code review — kimmonismus · 2026-11-08
- Agentic Creativity course recordings now publicly listed on YouTube — round · 2026-11-08
- LittlePly: a chess agent combining transformer policy, value search and bounded Stockfish help — andreisavu · 2026-11-08