Your LLM can't watch a screen recording. Screex turns one into text it can read.

Screex

Screen-recording understanding for agents. Screex turns a screencast into a queryable index of UI states — each with the on-screen text (OCR), what text changed since the previous state, a thumbnail, and a full-resolution keyframe — so an LLM/agent can produce an action transcript, answer questions, or generate a how-to guide / bug report from a recording.

• Training-free & model-agnostic — no fine-tuned UI model; any LLM can read the index.
• pip install-only — OCR via rapidocr-onnxruntime, no system binaries.
• Server-friendly runtime — uses headless OpenCV, so CI and Linux servers do not need GUI libraries just to build indexes.
• Cheap by design — the on-screen text is plain text (nearly free to read); full-res keyframes are escalated to only when the text is insufficient.
• ~70% lower token cost — in our GUI video-QA benchmarks, handing an agent the Screex index instead of raw video frames cut the input tokens sent to the model by around 70%, with little loss in answer accuracy.
• Fast OCR — tuned onnxruntime threading makes text extraction ~3.85× faster than the default.
• Narration-aware — with pip install 'screex[audio]', the index includes a timestamped transcript of the spoken audio, interleaved into the step transcript.

Good for: bug repros → reproduction reports · demos & Loom videos → how-to docs · tutorials → step lists · "what did the user do / what URL did they open?" Q&A over a recording.

Best on screen recordings. Screex is tuned for screencasts — mostly-static UI punctuated by discrete changes (clicks, typing, navigation). On that input it segments into a handful of meaningful states quickly. For general / continuous-motion video (camera footage, gameplay, talking-head clips) the change detector fires on nearly every frame, so prefer --fast with a higher --change-threshold (e.g. --fast --change-threshold 0.10) to avoid over-segmentation.

Example

A short screen recording of a login → settings → error flow becomes a timestamped step list:

screex transcript bug-repro.mp4 -o steps.md

steps.md:

# Transcript — bug-repro.mp4  (0:06)

## 0:00–0:01  ·  State 1
Acme Console · Sign in · Email: rushi@acme.io
**Appeared:** Acme Console, Sign in

## 0:01–0:02  ·  State 2
Dashboard · Welcome back, Rushi · Projects: 3
**Appeared:** Dashboard, Welcome back, Rushi
**Gone:** Acme Console, Sign in

## 0:03–0:04  ·  State 3
Settings > API Keys · New key: sk-live-9f2a · [ Save ]
**Appeared:** Settings > API Keys, New key: sk-live-9f2a

## 0:04–0:06  ·  State 4
Error: invalid API key format · Expected prefix 'sk_' not 'sk-'
**Appeared:** Error: invalid API key format

Prefer richer output? Hand the index.json to Claude via the bundled skill and ask for a bug report, a how-to guide, or answers to questions about the recording.

Install

From PyPI

pip install screex

For spoken-word narration in the index, also install the audio extra: pip install 'screex[audio]'.

From source

git clone https://github.com/blueprintparadise/Screex.git
cd Screex
pip install -e .          # add ".[test]" to also install pytest

Both give you a screex command (entry point screex.cli:main). Requires Python ≥ 3.9. The OCR models ship inside the rapidocr-onnxruntime dependency, so no separate download is needed.

Quickstart (CLI)

# Build the index for a screen recording
screex index path/to/recording.mp4 --fps 2
#   (or, without installing the package:)
python -m screex.cli index path/to/recording.mp4 --fps 2

This writes:

path/to/recording.screex/
  index.json            # the ScreenIndex (ordered UI states)
  frames/00000.png      # full-res keyframe per state
  frames/00000_thumb.png# thumbnail per state
  ...

index options

Transcript (no LLM needed)

Turn a recording straight into a timestamped markdown step list:

screex transcript path/to/recording.mp4 -o steps.md    # omit -o to print to stdout
screex transcript path/to/recording.mp4 --from-index path/to/recording.screex/index.json

By default index/transcript segment by on-screen text change, so a dialog or a status line appearing becomes its own step. Use --fast for motion-only segmentation on simple clips.

Need a different output format? screex transcript … --format {md,json,srt,vtt} emits markdown (default), the compact JSON index, or SubRip/WebVTT caption cues (one cue per UI state).

Inspect & search an index

Once you have an index.json, query it without writing any code:

screex info path/to/recording.screex/index.json          # states, duration, event histogram, warnings
screex search path/to/recording.screex/index.json "error"  # states whose text matches
screex search …/index.json --event click --since 5 --until 20   # by event type / time window

Both accept --json for machine-readable output.

What index.json contains

A schema_version, the source video/duration/sampled_fps, and an ordered list of states, each with: t_start / t_end, ocr_text (on-screen text lines), text_added / text_removed (text that appeared/disappeared vs the previous state — the strongest signal of what the user did), thumbnail / keyframe paths, and optional warnings for recoverable diagnostics such as OCR failures on individual frames. With --boxes, each state also carries boxes (per-line {text, box:[x,y,w,h]}); with --interactions, each state may carry interactions ({t, x, y, label} estimated cursor hotspots). With --events, each state (after the first) may carry an event ({type, t, region:[x,y,w,h], confidence, label?, field?, value?}) — a typed, region-grounded estimate of what the user did, the strongest "what happened" signal.

Privacy / redaction

Screex indexes whatever is on screen — for bug repros that often includes passwords, API keys, tokens, and PII. Pass --redact to mask those in the text/narration and blur the matching regions in the keyframe images before anything is written, so the .screex directory is safe to share. Detection is best-effort (regex + high-entropy tokens), not a compliance guarantee — review sensitive recordings before sharing.

Use as a Claude skill

Screex ships a SKILL.md that teaches Claude to build the index and turn it into one of three views: an action transcript, Q&A over the recording, or a how-to / bug report.

1. Install the package so python -m screex.cli is available in the environment Claude uses (pip install -e .).

2. Install the skill — the package bundles SKILL.md, so one command installs it where Claude Code discovers skills:

   screex skill --install                                          # ~/.claude/skills/screex/
   screex skill --install --dir <project>/.claude/skills/screex    # per-project
   screex skill --path                                             # just print the target path
   screex skill --check                                            # is the installed skill current?
   

3. Use it — in Claude Code, just ask in natural language, e.g.:

   • "Use screex to turn ~/Downloads/bug-repro.mp4 into a bug report."
   • "What steps does this screen recording show?"
   • "From this demo, write a how-to doc."

   Claude runs screex index, reads index.json, skims the on-screen text across states, and escalates to a full-res keyframe only when the text isn't enough — then produces the transcript / answer / document.

The skill is model-agnostic: the same index.json can be read by any LLM/agent, not only Claude.

Staying current: upgrading the package does not re-copy SKILL.md. After an upgrade run pip install -U screex && screex skill --install; screex skill --check tells you if your installed skill is behind the package.

Use as an MCP server

Prefer the Model Context Protocol? Screex ships an optional stdio MCP server exposing build_index, info, search, and transcript as tools to any MCP client:

pip install 'screex[mcp]'
screex mcp

How it works

recording → sample frames → segment into UI states → per state: OCR text + text-diff
          → write thumbnail + full-res keyframe → index.json
                                                      ↓
            views (agent-driven): transcript · Q&A · how-to / bug report

screex/core/:

• source — decode & sample frames (OpenCV)
• segment — group frames into settled UI states by visual change
• ocr — RapidOCR text extraction + text-diff between states
• index — the ScreenState / ScreenIndex schema (JSON)

screex/cli.py wires them into the screex index command.

Development

pip install -e ".[dev]"     # ruff + mypy + pytest + mss
ruff check screex tests
python -m mypy screex
python -m pytest -q

See CONTRIBUTING.md for the full contributor guide. On Windows, if you install the audio extra, set KMP_DUPLICATE_LIB_OK=TRUE before running the tests to avoid an OpenMP runtime conflict between PyTorch (faster-whisper) and ONNX Runtime (RapidOCR).

License

MIT © 2026 Rushikesh Hiray