io.github.LastSearch-HQPublished by GitHub account LastSearch-HQv1.4.3Updated Nov 7, 2026
Evidence-backed web research for AI agents with citations and confidence scores.
LastSearch gives an assistant evidence-backed web research: cited answers, per-claim verification, confidence scores and persistent research sessions.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Follow the launch instructions in the repository.
Reliable research infrastructure for AI agents. The research layer your agents are missing.
MCP server with real-time web search, evidence extraction, and structured citations. Drop into Claude Desktop, Cursor, Windsurf, LangChain, CrewAI, or any agent pipeline.
Instead of letting your AI hallucinate, lastsearch gives it real-time access to the web with structured, cited answers:
Every answer includes:
"fast" (default), "thorough" (auto-retry with rephrased queries), "deep" (premium multi-step agentic research: iterative think-search-extract-evaluate cycles with gap analysis, up to 4 steps, targets 0.85 confidence — requires LastSearch key + sign-in, 3x quota cost, falls back to thorough when exhausted)LASTSEARCH_API_KEY)Users with a LastSearch API key (ls_xxx) get enhanced verification:
Free LastSearch key users get a generous daily quota (100 premium queries/day, or ~33 deep queries/day at 3x cost each). When exceeded, queries gracefully fall back to keyword verification (deep falls back to thorough). Quota resets every 24 hours.
Sign up at lastsearch.ai for a free API key — 100 premium queries/day with full verification pipeline.
This auto-configures Claude Desktop. You'll need a LastSearch API key — get one free at lastsearch.ai/dashboard.
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
Add to your MCP settings:
Run as an HTTP server for browser-based clients, Smithery, or any HTTP-capable agent:
Every request needs a bearer token. By default a random token is generated at each launch and printed once to stderr; set LASTSEARCH_HTTP_TOKEN (16+ characters) to use your own:
The server exposes:
POST /mcp — MCP Streamable HTTP endpointGET /health — Health checkBy default the server only accepts requests addressed to 127.0.0.1, localhost or ::1, and refuses requests from browser pages on other origins. To listen on another interface, pass --host <address> explicitly (a warning is printed); the token is still required. Request bodies are limited to 1 MB.
Inside a container the server must listen on 0.0.0.0, so pass --host 0.0.0.0 explicitly, set a token, and publish the port on the host's loopback interface only:
Clients send Authorization: Bearer <LASTSEARCH_HTTP_TOKEN> with every request.
Note: All tools require a LastSearch API key (
ls_xxx). Get a free one at lastsearch.ai/dashboard.
Quick lookup:
"Use answer to explain what causes aurora borealis"
Higher accuracy:
"Use answer with depth thorough to research quantum computing"
Deep research (multi-step, requires LastSearch key):
"Use answer with depth deep to compare CRISPR approaches for sickle cell disease"
Deep mode runs iterative think-search-extract-evaluate cycles: gap analysis identifies missing info, follow-up queries fill the gaps, and claims/sources are merged across steps with final re-verification. Targets 0.85 confidence across up to 4 steps. Falls back to thorough without a LastSearch key or when quota is exhausted.
Contradiction detection:
"Use answer with depth thorough to check if coffee is good for health, and show me any contradictions"
Research session:
"Create a session called quantum-research, then ask about quantum entanglement, then ask how entanglement is used in computing"
Enterprise search:
"Use answer to search our Elasticsearch at https://es.company.com/kb/_search for our refund policy"
All API calls include automatic retry with exponential backoff on transient failures (429 rate limits, 5xx server errors). Auth errors fail immediately — no wasted retries.
Pre-built skills that teach coding agents when to use LastSearch tools:
Skills work with Claude Code, Codex CLI, Gemini CLI, Cursor, and more. View skills →
Apache 2.0
This package sends your requests (queries, URLs and documents you ask it to verify) and your LastSearch API key to the hosted LastSearch API at lastsearch.ai. Queries and results are stored there and used, in aggregate and without identifiers, to improve verification, as described in the Privacy Policy. The package itself collects no telemetry. Use of the hosted API is subject to the Terms.
LastSearch is a free, non-commercial passion project run by an individual developer. Answers, citations and confidence scores are AI-generated research, not professional advice; verify anything important against primary sources.
Source: apps/mcp/README.md at commit 4d25e7c
Fetch the signed evidence record for a previous answer (pass the result's shareId). Returns a typed record (retrieval → verification → decision) with per-source content hashes, the exact canonical bytes that were signed, their SHA-256, and an Ed25519 signature verifiable against GET /browse/evidence/public-key. Attests what LastSearch retrieved, verified, and returned — not what your agent did with it; not a compliance guarantee.
resultIdrequiredSearch the web for information on a topic. Returns URLs, titles, snippets, and relevance scores.
limitqueryrequiredFetch and parse a web page into clean text using Readability. Strips ads, nav, and boilerplate.
urlrequiredExtract structured knowledge (claims + sources + confidence) from a single web page using AI.
urlrequiredqueryFull deep research pipeline: search the web, fetch pages, extract claims, build evidence graph, and generate a structured answer with citations and confidence score. Use depth='thorough' for auto-retry with rephrased queries when confidence is low. Use depth='deep' for multi-step agentic research that identifies knowledge gaps and runs follow-up searches. On thorough/deep, every cited URL is probed for liveness: each source carries a urlHealth field (live | stale | dead | unchecked) and the result includes a citationHealth summary, so you can spot broken or likely fabricated links (checks fail open to 'unchecked'). Every fetched page is also scanned for hidden / non-rendered content (where most real-world prompt injections live) before extraction: hidden content is stripped, instruction-shaped text is counted, and each source carries an integrity signal plus a result-level contentRisk summary — a signal for your policy layer, not a guarantee. Instruction-directed spans are redacted from page text before extraction; pass strict=true to drop high-risk sources entirely. Every stored result has a signed evidence record — call the evidence tool with the result's shareId. Enterprise: use searchProvider to search internal data instead of the public web. DISCLAIMER: Results are AI-generated for informational purposes only — not financial, medical, or legal advice. Confidence scores are algorithmic estimates, not accuracy guarantees. Always verify critical information from primary sources.
depthqueryrequiredstrictsearchProviderCompare a raw LLM answer (no sources) vs an evidence-backed answer. Shows the difference between hallucination-prone and grounded responses.
queryrequiredFact-check an entire document (AI-generated report, competitive analysis, market research, news article). Pass either raw text OR a URL — we'll extract every atomic claim and verify each against live web sources. Returns per-claim verification status with sources, NLI scores, and an overall A-F grade. Use for auditing the accuracy of long-form content from other AI agents.
urltextdepthtitlemaxClaimsClarity — anti-hallucination answer engine. Three modes: (1) mode='prompt': Returns only enhanced system + user prompts with anti-hallucination techniques. No LLM call, no internet. Use this when you want YOUR OWN LLM (e.g. Claude) to answer using the enhanced prompts. (2) mode='answer' (default): Rewrites prompt, calls LLM with grounding instructions, returns answer with extracted claims. Fast, no internet. (3) mode='verified': Does #2, then runs full browse pipeline (search + extract + verify), fuses best of both — source-backed claims, evidence-based confidence. Use for maximum accuracy.
modedepthintentpromptrequiredverifycontextCreate a new research session. Sessions persist knowledge across multiple queries — each query builds on prior research.
namerequiredResearch a question within a session. Recalls prior knowledge, runs the research pipeline, and stores new claims. Later queries in the same session benefit from accumulated knowledge.
depthqueryrequiredsession_idrequiredQuery accumulated knowledge from a session without making new web searches. Returns previously verified claims relevant to the query.
limitqueryrequiredsession_idrequiredShare a research session publicly. Returns a shareable URL that anyone can view — great for sharing research findings with teammates, in reports, or on social media.
session_idrequiredExport all knowledge from a research session. Returns all verified claims, sources, and confidence scores accumulated across queries.
limitsession_idrequiredFork a shared research session to continue building on someone else's research. Creates a copy of all knowledge in your own session.
share_idrequiredSubmit feedback on a search result to improve future accuracy. Helps the self-learning engine tune verification thresholds.
ratingrequiredresult_idrequiredclaim_index90% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
Your question → Web search → Neural rerank → Fetch pages → Extract claims → Verify → Cited answer (streamed)npx lastsearch setup{ "mcpServers": { "lastsearch": { "command": "npx", "args": ["-y", "[email protected]"], "env": { "LASTSEARCH_API_KEY": "ls_xxx" } } }}{ "lastsearch": { "command": "npx", "args": ["-y", "[email protected]"], "env": { "LASTSEARCH_API_KEY": "ls_xxx" } }}# Start with HTTP transport (listens on 127.0.0.1 only)npx lastsearch --http
# Or set the port via environment variableMCP_HTTP_PORT=3100 npx lastsearch --httpexport LASTSEARCH_HTTP_TOKEN="$(openssl rand -base64 32)"npx lastsearch --http
curl -H "Authorization: Bearer $LASTSEARCH_HTTP_TOKEN" http://127.0.0.1:3100/healthdocker build -t lastsearch ./apps/mcpdocker run -p 127.0.0.1:3100:3100 \ -e LASTSEARCH_API_KEY=ls_xxx \ -e LASTSEARCH_HTTP_TOKEN="$(openssl rand -base64 32)" \ lastsearch --host 0.0.0.0{ "answer": "Aurora borealis occurs when charged particles from the Sun...", "confidence": 0.92, "claims": [ { "claim": "Aurora borealis is caused by solar wind particles...", "sources": ["https://en.wikipedia.org/wiki/Aurora"], "verified": true, "verificationScore": 0.82, "consensusLevel": "strong" } ], "sources": [ { "url": "https://en.wikipedia.org/wiki/Aurora", "title": "Aurora - Wikipedia", "domain": "en.wikipedia.org", "quote": "An aurora is a natural light display...", "verified": true, "authority": 0.70 } ], "contradictions": [], "reasoningSteps": []}npx skills add lastsearch-hq/lastsearch-skills