WebVector 0.1.0

First public release of WebVector — web research for AI agents in one tool call: search → read the full pages → rank → cited passages.

Packages (npm)

Package	Install	What it is
webvector	npm i webvector	the library — new WebVector().research(query) + AI SDK / Anthropic / OpenAI / LangChain bindings
webvector-mcp	npx -y webvector-mcp	MCP server: web_research, web_fetch, web_search, webvector_status (stdio + Streamable HTTP)
webvector-cli	npx -y webvector-cli search "…"	CLI: search, fetch, serp, doctor, init, config, mcp

Try it in 30 seconds

npx -y webvector-cli search "what changed in the MCP spec in 2026?" --stats
claude mcp add webvector -- npx -y webvector-mcp

No API keys required. ~12 MB install (lexical BM25 tier); add @huggingface/transformers or any embedding API key and it switches to on-device / hosted semantic search automatically.

Highlights

• 11 search providers (DuckDuckGo keyless, Brave, Serper, SerpAPI, Google CSE, SearXNG, Tavily incl. keyless, Exa, Perplexity, Wikipedia, custom) with automatic fallback chains
• Embeddings: none (BM25), local Transformers.js presets, OpenAI, OpenAI-compatible, Gemini, Voyage, Cohere, Mistral, Jina, Ollama, Vercel AI SDK models
• Stores: in-memory (sessions with TTL), Chroma, Qdrant, pgvector · Rerankers: local cross-encoder, Cohere, Voyage, Jina, LLM
• Retrieval: multi-query expansion, BM25 + vector hybrid (weighted RRF), near-duplicate removal, per-source caps, MMR
• Ingestion: polite concurrent fetcher (robots.txt, per-host limits, size/redirect/time caps), SSRF guard enforced at connect time, Readability→Markdown, PDF, charset sniffing, markdown-aware chunker with breadcrumbs and offsets
• Every passage carries URL, title, offsets and score; failures are reported per URL, never thrown

Docs: README · Configuration · Providers · Security