MikkoParkkola/nab

Token-optimized web fetcher + multilingual ASR + URL watcher. MCP 2025-11-25 and 2026-07-28 (partial: a call that does not wait). Rust. macOS arm64 first, cross-platform.
nab is a single Rust binary that does three things very well: it **fetches** any URL as clean markdown (with your real browser cookies and anti-bot evasion), it **analyzes** any audio or video file with on-device multilingual ASR and speaker diarization, and it **watches** any URL for changes and pushes notifications when content moves. Everything runs locally. There are no API keys to set up by default. The output is shaped for LLM context windows.
Why nab
- **Token-lean by design.** nab returns only what an LLM actually needs — clean markdown, BM25-lite query-focused extraction, and structure-aware token budgets — cutting the token cost of web research instead of dumping raw HTML into your context window.
- **Multimodal, fully on-device.** Transcribe and diarize any audio or video (FluidAudio / Parakeet TDT v3 on the Apple Neural Engine — 131× realtime on a 2-hour clip, 25 EU languages, word-level timestamps, optional Qwen3-ASR for zh/ja/ko/vi) and OCR images via Apple Vision (15 languages, ~10–50 ms). No cloud, no API keys.
- **Authenticated reach.** Real browser cookies, 1Password auto-login with TOTP/MFA, WebAuthn passkeys, fingerprint spoofing and WAF evasion — reach internal dashboards, SaaS apps, and paywalled research with the same command as a public URL.
- **Watch the web.** Subscribe to any URL via MCP resources — conditional GETs, semantic diff, adaptive backoff. RSS for the entire web.
- **Prompt-injection defense, on by default.** Hidden instructions addressed to your AI are surfaced to you, not silently executed by your model — see Security.
Everything is a single local Rust binary. No cloud backend, no API keys by default, output shaped for LLM context windows.
Quick start
**Tell your AI assistant** (recommended):
> Read https://github.com/MikkoParkkola/nab and install nab as my web fetching and audio analysis MCP server
Your agent will install the binary, wire itself up, and start fetching. Works in Claude Code, Cursor, Windsurf, and any AI with terminal access.
**Or install and try manually:**
Features
| Command | What it does |
|---------|--------------|
| nab fetch | Fetch any URL as clean markdown. HTTP/3, browser cookie injection (Brave / Chrome / Firefox / Safari / Edge / Dia), 1Password auto-login, fingerprint spoofing, fetch-time YARA-X redaction for prompt-injection/exfil signatures, 12 site providers. MCP fetch also supports query-focused extraction, readability, and token budgets. |
| nab browser | Explicit opt-in browser rendering for JS-heavy pages through a configured Chrome DevTools Protocol WebSocket endpoint. No Chromium is bundled and default nab fetch never auto-launches a browser or remote provider. |
| nab analyze | Transcribe and diarize. FluidAudio (Parakeet TDT v3) on Apple Neural Engine, 131x realtime on a 2-hour clip, word-level timestamps, 25 EU languages, optional Qwen3-ASR for zh/ja/ko/vi, optional active reading via MCP sampling. |
| nab watch add | Monitor a URL and push notifications via subscribable MCP resources. RSS for the entire web. Conditional GETs, semantic diff, adaptive backoff. |
| nab models fetch | Persistent install of inference model binaries. Supports fluidaudio (default on macOS Apple Silicon), sherpa-onnx (cross-platform Parakeet TDT, ~30× realtime CPU), and whisper (universal fallback, whisper-large-v3-turbo, 99 langs). |
| nab-mcp | MCP 2025-11-25 and 2026-07-28 server. stdio + Streamable HTTP. 12 tools, 4 prompts, 2+N resources, structured logging, sampling, roots, elicitation. |
| nab::content::ocr | Apple Vision OCR engine. 15 languages. Apple Neural Engine accelerated. ~10-50 ms per image. macOS only. |
Security: prompt-injection defense
Web pages increasingly carry instructions written **for the AI, not for you** — concealed in HTML comments, display:none / aria-hidden text, data-ai / data-mcp / data-agent attribute payloads, or WebMCP manifests. Fetch such a page with a naive tool and those hidden instructions land straight in your model's context, where they can be acted on. This is the prompt-injection-as-phishing class of attack.
nab treats every fetched page as hostile input and runs two local, non-networked guards **before any content reaches your agent** — on by default, no flag, no setup:
- **Secure Ingestion guard** — detects and strips machine-targeted markup that is invisible to humans (AI-addressed comments, hidden display:none / aria-hidden text, agent-only data-* payloads, WebMCP advertisements) and **reports** each detection at Info / Warn / Block severity, so you see what a page *tried* to tell your agent instead of it being silently executed.