neurolearn
nekith78/neurolearn
YouTube/Instagram/TikTok/local audio & video → text via 8 interchangeable backends (offline Whisper, YouTube subtitles, Groq, Gemini, OpenAI, Deepgram, AssemblyAI, custom). Plus topic research, channel subscriptions, PDF reports, and analysis. Claude Code skill, slash command, or standalone CLI.
Instalação
npx convez-ai nekith78/neurolearn1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
NEUROLEARN
Learn from videos. Universal transcription + analysis for YouTube, Instagram, TikTok, and local files.
What it does
Pass a video URL or local file → get a clean transcript (.txt with optional
timestamps + .srt). Pick from 8 interchangeable backends — fully offline
Whisper, instant YouTube subtitles, or fast cloud APIs. Built to be driven
from the terminal and from a Claude Code chat.
Highlights
- 8 backends, one interface —
subtitles·whisper-local·gemini·groq·openai·deepgram·assemblyai·custom. Switch with a single flag. - Smart cascade by default — tries YouTube subtitles first, falls back to Groq Whisper, then local Whisper.
- Any-length videos, transparently — Opus 24k recompression + adaptive silence-aligned chunking handles 4-hour videos on Groq's free tier. Silence-edge trim + word-variety hallucination filter prevent Whisper from inventing text on silent intros/outros — validated across music / tech-talk / interview / news / tutorial formats with zero false positives.
- Fast — 1 hour of audio → 3.2 minutes wall time on Groq (~19× realtime). See benchmarks.
- Research + subscribes — discover videos by topic across YouTube, IG, TikTok; subscribe to channels with incremental updates (YouTube pulls the
/videos+/shortstabs, so livestreams never masquerade as the latest upload). - Visual mode — keyframes + per-moment vision-LLM annotations. Inside Claude Code, neurolearn writes a manifest and Claude reads the frames natively — no extra API call.
- PDF reports —
neurolearn report --latestrenders structured PDFs (tutorial / vlog / generic layouts) from any batch. - Offline-capable —
--backend whisper-localis fully offline.--backend subtitlesonly talks to YouTube.
Why it exists
[!TIP] The killer use case: comprehensive research that combines Claude's web research with neurolearn-fetched video transcripts. Claude reads web sources; neurolearn fetches what was only said out loud in podcasts, conference talks, interviews, and long-form YouTube. Together they cover both the documented and the candid layer of any topic.
A concrete prompt that triggers the pattern:
"Research the current state of agentic AI safety. Use both
web sources and neurolearn to pull recent video interviews."
Claude reads web sources, runs neurolearn research "agentic AI safety" --days 90 --languages en --limit 20 in parallel, gets transcripts of 20
relevant talks and interviews, and writes a synthesis that catches what's
only said out loud alongside what's been written down.
Other patterns: daily channel digests, long-form video → structured PDF, multi-language research without speaking the languages, offline / private transcription. See docs/USE_CASES.md for full walkthroughs.
How smart cascade works
The default --backend smart flow — no flags needed:
flowchart LR
URL([Video URL]) --> YT{YouTube?}
YT -->|yes| SUB[Try YouTube subtitles<br/>~3s, free, no API key]
YT -->|no| DL
SUB -->|hit| OUT([Transcript])
SUB -->|miss| DL[Download audio]
DL --> RECOMP[Opus 24k mono recompress<br/>v0.14.1]
RECOMP --> SIZE{Fits Groq<br/>tier limit?}
SIZE -->|yes| TRIM[Silence-edge trim<br/>v0.15.1]
SIZE -->|no| CHUNK[Split at silence boundaries<br/>track per-chunk offset]
CHUNK --> TRIM
TRIM --> GROQ[Groq Whisper<br/>large-v3-turbo]
GROQ --> FILTER[Hallucination filter<br/>density + word-variety + blocklist]
FILTER --> OUT
GROQ -.fails.-> LOCAL[whisper-local fallback]
LOCAL --> OUT
End-to-end timestamps are preserved across chunking + silence-trim, so the final .srt matches the original video's timeline.
Install
The two paths most people want:
As a Claude Code plugin (recommended)
/plugin marketplace add nekith78/neurolearn
/plugin install neurolearn@neurolearn
uv sync && neurolearn config wizard
As a standalone CLI
uv tool install git+https://github.com/nekith78/neurolearn
neurolearn config wizard
For other paths (manual clone, skill folder, pip), system requirements, ffmpeg setup, and optional extras → see docs/INSTALL.md.
[!IMPORTANT] Cookies are file-only. YouTube/Instagram/TikTok cookies are registered via
neurolearn config set-cookies --from-file <path>(orsubscribes cookies set <platform> --from-file <path>). We deliberately do not support--cookies-from-browser— that yt-dlp flag pulls every cookie from your browser store into process memory. Heavy research without registered cookies WILL hit IP blocks. See docs/UNLIMITED_RESEARCH.md.
Quick start
# Transcribe a YouTube video — picks the best backend automatically
neurolearn transcribe https://youtu.be/dQw4w9WgXcQ
# Local file in Russian
neurolearn transcribe ~/Downloads/lecture.mp4 --language ru
# Whole channel, 10 most recent videos
neurolearn batch https://youtube.com/@anthropicai --limit 10
# Discover videos on a topic + auto-analyze with Gemini
neurolearn research "AI agents 2026" --days 14 \
--prompt "Compare design choices" --analyze-backend gemini
# Inside Claude Code
/transcribe https://youtu.be/xyz
From any Claude Code chat:
"Pull the latest 10 videos from @anthropicai via subtitles and write a topic summary"
Claude picks the right neurolearn subcommand, runs it, reads the output, and writes the summary. The skill produces transcripts; analysis is the LLM's job.
Documentation
| 💡 Use cases | The flagship Claude+neurolearn research pattern, plus 7 supporting workflows: channel digests, long-form → PDF, multi-language research, offline transcription, dataset building, claim verification, quote mining. |
| ⏱ Benchmarks | Wall-clock times for every feature on a 3:30 reference video, normalized to per-hour-of-audio rates. |
| 📖 Usage | Every command, every flag, with examples. transcribe · batch · analyze · research · subscribes · report · history · triggers · config |
| ⚙️ Backends | The 8 transcription backends compared. Hardware guide. Whisper model comparison. Groq size + hallucination handling. Privacy table. |
| 🤖 Claude Code integration | Plugin install, slash commands, switching backends from chat, secure key handoff, visual mode in Claude Code. |
| 💻 Install details | Per-OS install, optional extras, first-run wizard, HF_TOKEN warning. |
| 🚀 Unlimited research | 3-layer anti-block stack: cookies → PO Token plugin → residential proxy. When YouTube rate-limits your IP. |
| 🩹 Troubleshooting | yt-dlp 403, missing API keys, CUDA errors, missing ffmpeg, distil model on Mac, long-video timeouts. |
| 🏗️ Architecture | For developers: Transcriber Protocol, smart cascade internals, adding a backend, cross-OS specifics. |
| 📜 Changelog | Version-by-version detail. |
| 🤖 Agent reference | Full CLI surface, exit codes, state semantics — for LLM agents driving the skill. |
| 🍪 Cookies walkthrough | Step-by-step screenshots for the yt-dlp 403 fix. |
License
MIT — see LICENSE.