FlowScope

Given one or more YouTube product-demo URLs, FlowScope downloads the video, transcribes it (captions or local Whisper), extracts and deduplicates representative screenshots of distinct screens, aligns them to the narration, and uses an LLM to produce a structured breakdown of the product's flow, screens, and UX patterns — plus cross-video comparison.

For personal competitive UX research use. Downloaded videos and derived analysis stay local.

Use it from an LLM (MCP)

mcp/ is an MCP server that exposes this pipeline as tools, plus two installable skills, so any MCP-capable client can run the analyses for you:

Compare the onboarding flows in these two videos: <url> and <url>

Start the backend, then install into your client:

cd backend && uvicorn app.main:app --port 8000     # terminal 1

# Claude Code: one step, brings the skills and the server registration
/plugin marketplace add DavidNgugi/flowscope
/plugin install flowscope@flowscope

For any other harness, generate the client config — the shapes genuinely differ (VS Code's own file uses servers where everyone else uses mcpServers, and Codex is TOML):

uvx --from flowscope-mcp flowscope-mcp-install --help   # every client and its path
uvx --from flowscope-mcp flowscope-mcp-install cursor   # or vscode, codex, gemini

(flowscope-mcp-install is a console script inside the flowscope-mcp distribution, hence --from. If you installed the package locally, the bare flowscope-mcp-install works too.)

The two bundled skills install separately, into any Agent Skills-compatible client. .agents/skills/ is the cross-client convention, so one copy serves most harnesses:

mkdir -p .agents/skills && cp -r mcp/skills/* .agents/skills/
npx skills add DavidNgugi/flowscope        # or via the skills CLI

Ask the agent to run flowscope_health_check — it reports whether the backend, ffmpeg, yt-dlp, and an LLM provider key are all in place, so a missing prerequisite is a clear message rather than a failed ten-minute job.

Then just describe what you want:

Tear down the onboarding UX in this demo: <youtube-url>

Compare the onboarding flows in these two demos. Where do they diverge, and what's the trade-off each one makes?

• <youtube-url-1>
• <youtube-url-2>

In this demo, how many steps does signup take, and what does each one ask for? <youtube-url>

Analysing a video takes 2–10 minutes and makes one vision call per distinct screen, so it costs real money. Results are cached by YouTube video id — asking "what has FlowScope already analysed?" is free, and re-reading a report costs nothing. The bundled skills teach the agent to check the cache before spending.

The server is a stateless client of the backend's REST API and imports none of the application code, so it installs without ffmpeg bindings or Whisper.

See mcp/README.md for the tool list, mcp/docs/install.md for the per-harness install reference, mcp/docs/usage.md for the full set of example prompts and troubleshooting, mcp/docs/skills.md for the skills matrix, and mcp/docs/publishing.md and mcp/docs/security.md for the details.

The MCP package releases on its own cadence, from version tags. Bumping is a single command because six declarations in five files must agree:

cd mcp && ./.venv/bin/python scripts/version.py bump 0.2.0

Then commit, tag v0.2.0, and push the tag — CI verifies everything, publishes to PyPI, publishes to the MCP Registry, and attaches the wheel to the GitHub release.

Setup

brew install ffmpeg

cd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # fill in a provider key (see Models below)
uvicorn app.main:app --reload --port 8000

cd frontend
npm install
npm run dev

Open http://localhost:5173.

Models

The LLM layer is provider-neutral: adding a vendor is configuration, not code. Four providers are built in — anthropic, openai, deepseek, and openai_compatible (any vendor or local server that speaks the OpenAI chat-completions format, e.g. xAI, Mistral, Groq, Together, OpenRouter, vLLM, Ollama).

Selection is by purpose, so frames and text can run on different models:

LLM_PROVIDER=openai            # default provider
LLM_MODEL=gpt-5-mini           # default model
# Optional per-purpose overrides (each falls back to the default):
# LLM_VISION_PROVIDER / LLM_VISION_MODEL          frame analysis (needs vision)
# LLM_SYNTHESIS_PROVIDER / LLM_SYNTHESIS_MODEL    one video's flow
# LLM_COMPARISON_PROVIDER / LLM_COMPARISON_MODEL  cross-video comparison

Per-provider credentials: ANTHROPIC_API_KEY, OPENAI_API_KEY, DEEPSEEK_API_KEY, or OPENAI_COMPATIBLE_API_KEY + OPENAI_COMPATIBLE_BASE_URL. A provider with no key is simply unavailable; selecting it produces a clear configuration error naming the variable to set.

Notes that save time:

• Vision is required for the frame stage. deepseek-flash accepts images; deepseek-v4-pro does not, and a text-only model produces an explicit error rather than silently analysing nothing.
• Reasoning models (gpt-5.x, o-series) spend part of the completion budget on hidden reasoning, so the adapter applies a floor and LLM_REASONING_EFFORT (default low). A budget sized for the visible answer alone fails with finish_reason=length and no tool call.
• Claude Sonnet 5.5 rejected forced tool use, so those models are routed to automatic tool choice; other Claude models keep forced tool use.
• Cost is recorded per call with the rate snapshotted onto the row. Models without a published rate in app/llm/catalog.py record tokens with a null cost rather than a guess.

Analysing frames that already exist

Downloading and frame extraction are separate from the paid stages, so a run that stopped at the LLM step — or a re-run against a different model — costs nothing to repeat:

python scripts/evaluate_existing_frames.py --data-dir ./data          # resumable
python scripts/export_analysis.py --data-dir ./data --out ./export    # plain files

Docker

docker compose build          # rebuild after any change under backend/app
docker compose up -d          # backend :8000, frontend :5180
open http://localhost:5180

backend/data is bind-mounted to /app/data, so the database, media and frame files live on the host: they survive rebuilds, can be inspected directly, and are shared with a local uvicorn run. (The older backend_data named volume is still declared but unused, so its contents are not lost.)

Because the app copy is baked into the image at build time, a code change needs docker compose build — up -d alone reuses the old image.

Data portability: frame paths are stored absolute, but the API resolves them by their path below the last media/ segment, so the same database works from the host and from inside the container.

YouTube authentication

YouTube may challenge yt-dlp with Sign in to confirm you're not a bot, even for public videos. FlowScope can reuse an authenticated YouTube session in one of two ways (configure only one):

• Local development: set YTDLP_COOKIES_FROM_BROWSER=chrome in backend/.env (also supported: firefox, safari, and other browsers recognized by yt-dlp), then restart the backend.
• Docker/server: export a Netscape-format youtube-cookies.txt, mount it into the backend container, and set YTDLP_COOKIE_FILE to its container path.

For Docker, add a read-only mount under the backend service, for example:

volumes:
  - backend_data:/app/data
  - ./secrets/youtube-cookies.txt:/app/secrets/youtube-cookies.txt

Then set YTDLP_COOKIE_FILE=/app/secrets/youtube-cookies.txt in backend/.env. Cookie files contain account credentials; keep them out of source control and refresh the export if YouTube invalidates the session.