# LastSearch [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/lastsearch-hq/lastsearch  
**GitHub Stars:** 20  
**npm Downloads (last month):** 553  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/lastsearch

## Description
Evidence-backed web research for AI agents with citations and confidence scores.

## Tools
Capabilities this server exposes over MCP:

- **search** — Search the web for information on any topic
- **open** — Fetch and parse a web page into clean text
- **extract** — Extract structured claims from a page
- **answer** — Full pipeline: search + extract + cite. `depth`: `"fast"`, `"thorough"`, or `"deep"
- **compare** — Compare raw LLM vs evidence-backed answer
- **clarity** — Anti-hallucination answer engine — three modes: prompt (prompts only), answer (LLM), verified (LLM + web fusion)
- **session_create** — Create a research session (persistent memory)
- **session_ask** — Research within a session (recalls prior knowledge)
- **session_recall** — Query session knowledge without new web search
- **session_share** — Share a session publicly (returns share URL)
- **session_knowledge** — Export all claims from a session
- **session_fork** — Fork a shared session to continue the research
- **feedback** — Submit feedback on a result to improve accuracy

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "lastsearch": {
    "command": "npx",
    "args": ["-y","lastsearch"]
  }
}
```

## Documentation

## What LastSearch MCP server does

LastSearch MCP server exposes an evidence-oriented web research workflow through MCP. Agents can search the web, retrieve readable page text, extract structured claims, and produce answers with citations and confidence information. Results can include claim verification status, verification scores, consensus counts, source authority, citation health, and detected contradictions.

The service is suited to questions where the agent must show where an answer came from. It supports both one-off lookups and longer investigations that build on earlier findings. The `answer` tool combines searching, extraction, and citation into one operation, while lower-level tools allow an agent to control each stage.

## How it works

A typical research path searches across providers, fetches relevant pages, extracts claims, and checks those claims against source text. Verification can use keyword matching, semantic retrieval, evidence reranking, source authority, cross-source agreement, contradiction detection, and quote matching. Premium features add semantic retrieval and reranking; without premium keys, the system falls back to keyword-only verification.

The `answer` tool accepts three depth modes. `fast` performs one search, extraction, and verification pass. `thorough` can repeat targeted retrieval for weak claims, run counter-query checks, and compare extraction passes. `deep` performs iterative research with gap analysis and follow-up queries, then re-verifies the combined result. Deep mode requires a LastSearch key and sign-in; without them, or when its quota is exhausted, it falls back to thorough mode.

## Tools and capabilities

Available MCP tools include:

- `search` for web queries and `open` for converting a page into clean text.
- `extract` for pulling structured claims from a page.
- `answer` for a complete cited research response with selectable depth.
- `compare` for comparing a raw language-model answer with an evidence-backed answer.
- `clarity` for prompt-only, language-model, or web-fused verified answers.
- `session_create`, `session_ask`, and `session_recall` for persistent research memory.
- `session_share`, `session_knowledge`, and `session_fork` for sharing, exporting, and continuing sessions.
- `feedback` for submitting result feedback to improve accuracy.

Streaming API responses can report trace events, discovered sources, answer tokens, the final result, and completion status. Deep responses can also expose research steps, including queries, gap analysis, claim counts, and confidence per step.

## Limitations and notes

Sessions require a LastSearch API key for identity and ownership. Deep research consumes more quota than ordinary queries: the README states that each deep query costs three times quota and that deep mode has a 100-query-per-day limit. Deep research targets a confidence threshold of 0.85 and can run up to four total research steps.

Confidence is evidence-based rather than a self-rating from the language model, but it should still be treated as a decision signal. Single-source claims can be marked weak, conflicting evidence is surfaced, and citation health can classify URLs as live, stale, or dead in thorough and deep modes. External calls retry transient rate-limit and server errors with backoff, while authentication errors fail immediately.

LastSearch MCP server does not eliminate the need to inspect sources for high-consequence work. The provided material describes API keys and web/API behavior, but does not specify a client installation procedure, MCP transport, or client-specific configuration.

_Full upstream README: https://allmcps.com/mcp/lastsearch/readme_