Watch Skill

Give any AI agent the ability to actually watch video.

Watch Skill turns any video — a URL from 1800+ sites, a live HLS/DASH stream, a local file, or a recording of the agent's own output — into a persistent, searchable index of frames, on-screen text, and transcript. Agents ask questions in seconds, get answers with auditable confidence and timestamp citations, and learn from their own mistakes. One engine, three surfaces: MCP (13 tools), CLI, and REST.

THE LOOP, live: iteration 0 flags TOTAL: $NaN as critical with a suggested fix → the agent fixes the code → iteration 1 verifies FIXED and renders this GIF.

60-second quickstart

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/oxbshw/watch-skill/main/scripts/install.sh | sh

# Windows
powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/oxbshw/watch-skill/main/scripts/install.ps1 | iex"

The installer bootstraps uv/Python if missing, self-heals the binary deps (ffmpeg, yt-dlp, deno via watch-skill doctor), and watch-skill setup writes the MCP config into every agent it finds — Claude Code, Claude Desktop, Cursor, Codex CLI, Windsurf, Gemini CLI — backing up anything it touches. Then:

watch-skill watch "https://youtu.be/..." "what happens in this video?"
watch-skill ask <video_id> "when exactly does the demo crash?"
watch-skill serve            # MCP stdio (--http for streamable HTTP)

Or skip the CLI entirely: restart your agent and say "watch this video: <any URL> — what happens at 0:10?". Manual per-agent config lives in docs/agents/.

Features

• Watch anything. 1800+ sites via yt-dlp (self-updating on extractor breakage), direct media URLs, HLS/DASH live streams (bounded capture), local files, and screen/window/browser recording.
• Frame budgets that respect your context window. Scene detection + perceptual-hash dedup spend the budget on distinct content: 512 px frames, hard cap of 100 per video, ≤2 fps, duration-tiered — with a dense focused mode for --start/--end windows.
• Offline by default. Platform captions (original language preferred over auto-translations) → local faster-whisper with RAM-aware model selection → cloud STT only if you opt in. The video file never leaves the machine — enforced by tests, not policy. Point vision at Ollama and the entire pipeline runs with zero cloud calls.
• Analyze once, ask forever. A schema-versioned SQLite index (FTS5 + local ONNX embeddings, hybrid retrieval) persists across sessions. Follow-ups answer in seconds without re-processing; search_videos spans every video ever watched. Vector scoring is numpy-batched: 122 ms over 10k×384 vectors vs 5.46 s pure-Python (45×).
• Answers you can trust (v0.6). Every answer carries a calibrated confidence score built from real retrieval signals (top-hit strength, margin over the runner-up, cross-kind evidence agreement, lexical anchoring). Below the bar, an escalation ladder runs cheapest-first — dense re-sampling → 2× zoom-crop re-OCR → a verify pass where the model is shown the exact frames it is about to cite. Still unsure? The honest floor says so plainly, with the closest real moments. Fabricated timestamps cannot survive composition (test-enforced).
• It learns from its mistakes (v0.6). report_mistake turns a correction into a classified lesson in ~/.watch-skill/lessons.db — local, never uploaded — injected into future similar questions across every agent on the machine. Every mistake becomes a replayable eval: watch-skill evals run reports the pass-rate over time.
• Spends tokens like they're yours (v0.6). Text-first answers (zero image tokens unless genuinely uncertain), a semantic answer cache (repeats are free, marked cached: true), a per-question token budget the ladder respects, and a savings meter. On this machine, 9 answers served ≈ 86,647 tokens saved vs raw-frame injection (watch-skill stats).
• Reads your language. Per-script OCR models auto-selected and auto-downloaded (Arabic, Cyrillic, Devanagari, Korean, …; benchmarked per script — Arabic at 100% char-hit on the bench render), Arabic hamza/diacritic-folded search, CJK substring matching, and a multilingual embedding model — ask in English about an Arabic transcript (en→ar 0.58 vs ~0.0 for distractors) and retrieval still lands.
• THE LOOP. The agent records its own output (browser, screen, window), gets a structured critique against natural-language pass criteria, fixes the code, and re-verifies — with a before/after proof GIF.
• Fast where it counts. Cold CLI start ~1.2 s; a full 10-second watch (scenes + frames + OCR + local whisper) in 32.9 s warm on an 8 GB-RAM, no-GPU machine; MCP/REST servers keep models resident so agent follow-ups skip load time. 284 tests, offline.

Works with your agent

Statuses are honestly graded — machine-tested ✅ (full end-to-end run in the agent), machine-configured ◐ (watch-skill setup wrote the config on a real machine and the server answered an MCP initialize; no in-app chat run), doc-verified ☑ (matches the agent's official docs, not executed here).

Full matrix with per-agent install, config, and 3-step smoke tests: docs/agents/README.md. An AGENTS.md adapter covers agents that read repo-level instructions.

Examples

Architecture

Thin surfaces, one core. src/watch_skill holds all logic; MCP, CLI, and REST are wrappers that never diverge.

flowchart LR
    subgraph agents["any agent"]
        CC[Claude Code] & CU[Cursor] & CX[Codex] & GA["...via REST"]
    end
    subgraph surfaces["surfaces (thin)"]
        MCP["MCP stdio/HTTP<br/>13 tools"] --- CLI[CLI] --- API[REST + OpenAPI]
    end
    subgraph core["src/watch_skill — all logic"]
        AC[acquire<br/>yt-dlp→fallbacks<br/>+ LRU cache] --> PE[perceive<br/>scenes · phash dedup · OCR]
        PE --> TR[transcribe<br/>captions → local whisper]
        TR --> IX[(index<br/>SQLite FTS5 + vectors<br/>+ answer cache)]
        PE --> VI[vision<br/>cheap/strong tiers<br/>5 providers]
        VI --> IX
        IX --> AN[answer<br/>confidence · escalation<br/>honest floor]
        LS[(lessons<br/>~/.watch-skill/lessons.db)] --> AN
        AN --> LS
        LP[THE LOOP<br/>capture → critic → diff] --> VI
    end
    agents --> surfaces --> core

Deep dive: docs/architecture.md — including "add a vision provider in ~20 lines" and "add a new Loop type".

How it compares to claude-video

Watch Skill began as an attempt to surpass claude-video — the skill that first gave Claude a video input, and the source of ideas we kept (token-aware frame budgets, captions-first transcription, focused mode). Credit where due: it is simpler to adopt for Claude-only workflows and has no engine to install. What's different:

| | claude-video | Watch Ski