How Watch Skill compares
September 8, 2026 · View on GitHub
Honest comparison against the other ways to give an agent video. Numbers were checked on 2026-08-08 and will drift; the linked repositories are the source of truth.
If you only want the short version: Watch Skill is heavier to install than a single-purpose script and much younger than the big screen-recording projects. What it gives you in exchange is a persistent index, evidence you can audit, and a verification loop. If you need "transcribe this one video" and nothing else, a smaller tool is a reasonable choice.
The options
| Watch Skill | claude-video | mcp-video-analyzer | screenpipe | Upload to a frontier model | |
|---|---|---|---|---|---|
| Stars | 267 | 14,528 | 42 | 20,816 | — |
| Last release | v1.4.3, 2026-09-06 | v0.2.0, 2026-06-29 | active | active | — |
| Last commit | 2026-08-09 | 2026-06-30 | 2026-08-04 | 2026-08-08 | — |
| Open PRs | 1 | 71 | — | — | — |
| Frames + OCR + transcript | yes | yes | yes | screen only | varies |
| Persistent index across sessions | yes | no | no | yes (screen) | no |
| Ask again without reprocessing | yes | no | no | yes | no |
| Timestamped evidence per answer | yes | no | no | partial | no |
| Cross-video synthesis | yes | no | no | no | no |
| Verify an agent's own work | THE LOOP | no | no | no | no |
| Corrections persist as lessons | yes | no | no | no | no |
| Runs fully offline | yes | partial | partial | yes | no |
| Surfaces | skills, MCP, CLI, REST, Python | Claude skill | MCP | MCP, app | API |
| Install size | ~200 MB standard / ~600 MB all | small | small | app | none |
Against claude-video
This is the closest comparison, so it deserves detail. claude-video does the same core job — download a video, pull frames, transcribe, hand it to Claude — and it earned 14.5k stars in three months by being genuinely easy to start.
Where the two differ:
It forgets. Watch Skill remembers. claude-video processes a video and hands the result to the model. Ask a second question tomorrow and it downloads and transcribes again. Watch Skill writes to a persistent index, so follow-ups are a lookup. On a 40-minute video that is the difference between seconds and minutes, every time.
Evidence. Watch Skill answers carry timestamps, a confidence value, and the frames and transcript lines they came from. You can check the answer against the video. That matters when the answer feeds a bug report.
Verification. THE LOOP records an agent's browser or desktop session, checks it against plain-language criteria, and produces before/after proof after a fix. No other tool in the table does this.
Maintenance. As of 2026-08-08 claude-video has 71 open pull requests and its last commit was 2026-06-30. It is one person's project and that person appears to be busy. Several of those open PRs fix problems Watch Skill does not have:
| Open there | Here |
|---|---|
| #112 add offline whisper | local faster-whisper is the default fallback |
| #116 non-English videos yield no transcript | original-language captions are preferred over auto-translations (acquire/ytdlp.py) |
| #113–#115 caption language and PO-token failures | a caption ladder with a documented fallback order |
| #107–#110 Windows encoding and permission errors | Windows is a first-class CI target |
#117, #122 ffmpeg 8+ removed -vsync | the video and audio pipelines never used it; the benchmark frame extractor did, and now passes -fps_mode:v passthrough instead (bench/video_backends/realmedia.py). watch-skill doctor names the 5.1 floor that option implies |
None of this is a criticism of its author. A project that gets popular faster than one person can maintain it is a normal outcome, and the code was good enough to earn those stars.
Where claude-video is better: it is far smaller and starts faster. If your need is "summarize this YouTube link" and you will never ask a second question, the extra machinery here is overhead you do not need.
Coming from it? See Migrating from claude-video.
Against screenpipe
Different category, often confused. screenpipe records your screen continuously and makes the history searchable. Watch Skill takes a video you point it at — a URL, a file, a stream, or a capture it makes on purpose — and builds auditable evidence from it.
Use screenpipe if you want an always-on record of your own machine. Use Watch Skill if you want an agent to watch a specific thing and prove what it saw. They coexist; screenpipe is not an MCP video-understanding tool and Watch Skill is not a 24/7 recorder.
Against uploading to a frontier model
Frontier models are very good at watching video and getting better. For a one-off question about a short clip, uploading is often the right answer.
The gaps that remain:
- Cost at repetition. Every question re-sends the video. Watch Skill processes once and answers from the index.
- No provenance. You get an answer, not the frame it came from.
- Length. Long recordings hit context limits. Scene-aware extraction spends a frame budget on distinct moments instead.
- Privacy. Some footage cannot leave the machine. An Ollama configuration keeps the whole pipeline local.
- Non-English. Script-aware OCR routing and original-language captions are work that general models still do unevenly.
Watch Skill uses frontier models for visual Q&A. The point is not to replace them but to stop paying for the same video twice.
Where Watch Skill is weakest
- Install weight. ~200 MB standard, ~600 MB with OCR, Whisper, and the browser. Smaller tools exist and are honestly smaller.
- Young. v1.0 shipped 2026-07-12. Fewer users have hit the edge cases.
- Vision still costs money unless you run Ollama locally.
- One maintainer right now. The same risk the table calls out elsewhere applies here; CONTRIBUTING.md is the honest answer.