What separates this from a search tool is everything after extraction: each claim is decomposed into atomic facts, matched back against the text of the pages it came from, cross-checked across independent sources, and penalised where contradictions appear — so verified, consensus level and source authority arrive as fields an agent can act on rather than as prose. Sessions are the other half: a shared session can be forked, which turns one agent's verified research into another's starting point.
A research pipeline exposed as tools: search, fetch pages, extract atomic claims, verify each one against the actual source text, and return an answer with citations and a confidence score. Sessions accumulate what has been verified, so later questions build on earlier ones.
- Web search returning URLs, titles, snippets and relevance scores
- A page fetched and reduced to clean text, with ads, navigation and boilerplate stripped out
- Structured knowledge extracted from a single page — claims with their sources and confidence
- The full pipeline in one call: search, fetch, extract, build the evidence graph, and answer with citations and a confidence score
- Three depth settings — a single pass for quick lookups, an iterative confidence-gated loop for fact-checking, and a multi-step mode with gap analysis for complex questions
- A whole document fact-checked: every atomic claim extracted and verified against live sources
- A raw model answer compared against an evidence-backed one, to show what grounding actually changes
- An anti-hallucination mode that can return only the enhanced prompts, with no model call made at all
- Research sessions created, asked within — recalling prior verified claims — recalled without new searches, exported as accumulated knowledge, shared as a public URL, and forked from someone else's share
- Feedback submitted on a result, or on one specific claim, which tunes the verification thresholds
- Per-claim fields an agent can branch on: verified, verification score, consensus count and level, source authority, and any contradictions detected
An API key, free to obtain, passed to the server as LASTSEARCH_API_KEY; the server itself runs over stdio through npx. The verification engine is a hosted service, so calls go out to lastsearch.ai even when the server runs locally — the MCP server and frontend are open source, the engine is not. Sessions need the key, since that is what establishes ownership of one. The free tier covers 100 premium queries a day, with deep queries costing three times as much, and when the quota runs out queries fall back to keyword-only verification rather than failing; it resets every 24 hours.
One command plus a key — npx lastsearch setup, then supply credentials