BrowserAI-Dev (repository name BrowseAI-HQ/BrowserAI-Dev; product data contains naming such as LastSearch-HQ/lastsearch, with package names npm: lastsearch, PyPI: lastsearch, LangChain: langchain-lastsearch; the old name remains usable and automatically redirects) is research infrastructure for AI agents, focused on "evidence-backed web search." The official description emphasizes: real-time web search, evidence extraction, verification, and structured citations; every claim is backed by a URL, every answer comes with a confidence score, and it is said to avoid hallucinations through a multi-step verification process. It offers three access methods: an MCP server, a REST API, and a Python SDK, along with entry points such as the website, Playground, API documentation, Alternatives, and Discord.
In terms of purpose, the project is positioned as "reliable research infrastructure for AI agents (Grounded Intelligence)," used to let agents obtain verifiable sources, confidence, and citations when answering questions, rather than relying only on the model's own knowledge. It also provides Clarity anti-hallucination capability and multiple depth modes, adapting to different needs from quick fact queries to deep multi-turn research.
In terms of main capabilities, the officially listed verification and confidence mechanisms include: atomic claim decomposition (splitting compound claims into independently verifiable facts); hybrid keyword and semantic retrieval (keyword matching + dense vector semantic matching with fused ranking, an advanced-tier capability, falling back to pure keyword when not configured); semantic evidence reranking (using a dedicated verification model trained on 1.4M+ claim-evidence pairs, with contradiction penalties and paraphrase bonuses); multi-provider parallel search to increase source diversity; domain authority scoring (covering 5 tiers, 10,000+ domains, and dynamically adjusted with real verification data); source citation verification; cross-source consensus (support from 3 or more independent domains is "strong consensus," while a single source is marked "weak"); contradiction detection (semantic conflict identification, presented in the response and penalizing confidence); multi-turn consistency (cross-checking across independent extraction rounds in thorough mode); automatically calibrated confidence (based on multiple factors such as verification rate, domain authority, number of sources, consensus, domain diversity, claim grounding, source freshness, and citation depth, and adjusted with user feedback); targeted follow-up search by claim; counter-query verification (using "whatadversarial query stress testing that "will falsify it"); confidence-gated iterative retrieval (verify → if weak claims remain → generate targeted queries → search → verify again, up to 3 rounds, with early stopping when queries repeat or the threshold is met).
In addition, there is real-time citation health checking (under thorough/deep, the server probes each cited URL and classifies it as resolves/stale/dead, used to identify link rot or fabricated citations), content integrity (before extracting the original text, scan the raw DOM, strip hidden/unrendered content and instruction-type text, count and annotate integrity and contentRisk; high-risk sources can be excluded via strict: true; the official documentation states that this rule produced no high-risk false positives on a 36-page open evaluation set, but emphasizes that this is only a small-sample count, not a guarantee, and adaptive attacks may still bypass it), signed evidence records (GET /browse/evidence/{shareId} returns a replayable typed record of retrieval → verification → decision, binding claims to sources and sources to page-text SHA-256 hashes, with an Ed25519 signature; the public key is at /browse/evidence/public-key; the SDK provides client.get_evidence(id) and LastSearch.verify_evidence(bundle, pem)). In the returned fields, claims include verified, verificationScore, consensusCount, consensusLevel; sources include verified, authority, urlHealth, integrity; at the top level there are also contradictions, citationHealth, and contentRisk summaries, making it easy for agents to make trust decisions programmatically.
In terms of usage, access methods include: MCP server (Claude Desktop, Cursor, Windsurf, can be manually written into the MCP configuration, requires setting the LASTSEARCH_API_KEY environment variable), REST API (Authorization: Bearer ls_xxx), Python SDK (api_key="ls_xxx", plus an AsyncLastSearch asynchronous interface), LangChain integration (provides LastSearchSearchTool, LastSearchAnswerTool (verified), LastSearchExtractTool, LastSearchCompareTool, LastSearchClarityTool — five tools). Depth levels are divided into fast, thorough, and deep; deep mode executes a "think-search-extract-evaluate" iterative loop, including gap analysis, targeted follow-up queries, cross-step merging of claims and sources with a final re-verification, a target confidence threshold of 0.85 (DEEP_CONFIDENCE_THRESHOLD), and a maximum of 3 follow-up steps (MAX_FOLLOW_UP_STEPS, 4 rounds total including the initial one). The deep response includes reasoningSteps; the streaming endpoint sends trace, sources, token, result, and done events via SSE. All external calls (search providers, LLM, page scraping) automatically retry with exponential backoff on 429/5xx, and fail immediately on 401/403. Research memory (Sessions) can accumulate verified claims across multiple queries, returns recalledClaims and newClaimsStored, and supports public sharing and forking, facilitating multi-agent collaboration. Complex queries are automatically decomposed into sub-queries with intent labels (definition, evidence, comparison, counterargument, technical, historical, etc.); simple factual queries skip planning to reduce latency.
In terms of applicable scenarios, it is suitable for AI agents that need traceable web research, retrieval-augmented applications, multi-agent collaborative research, and automated workflows that need to make trust decisions programmatically based on confidence, consensus, contradictions, and citation health; enterprise scenarios support using Elasticsearch, Confluence, and custom endpoints to replace or supplement public web search, with optional zero data retention.
In terms of limitations: the verification engine is a hosted service and does not run locally (the MCP server and frontend are open source and can run locally); all API access requires an ls_xxx key; the free tier allows 100 advanced queries per day (deep counts as 3x, about 33 per day), and after the quota is exhausted it degrades to keyword verification (deep falls back to thorough), while Pro can remove the limits; when no advanced key is set, only keyword verification runs, and semantic retrieval and reranking are unavailable; the official content integrity filtering explicitly states that it is only a policy-layer signal rather than a guarantee, and cannot defend against adaptive attacks; evidenceThe signature only proves what the service retrieved, verified, and returned, and when, and is not a compliance guarantee. Information such as pricing, specific prices, details of supported LLMs, and enterprise plan prices is not provided in the current materials.
