setup-video-vision

name	setup-video-vision
description	Use when the user says "set up video vision", "install video vision", "configure video transcription", asks to configure agent-video-vision, or hits first-run setup errors (missing engine, venv, or Parakeet model). Interactive wizard — choose a transcription engine, install deps, download models, verify.

Setup Video Vision

Guide the user through configuring agent-video-vision step by step. Ask one question at a time using multiple choice. Everything runs locally — there is no API key or cloud backend to configure.

Step 0: Build the server (fresh installs)

A fresh plugin install ships source only — the MCP server must be built once. Find the repo root (it is two directories up from this SKILL.md file; on Claude Code plugin installs it lives under ~/.claude/plugins/). Then check whether mcp-server/dist/index.js exists there. If it does not:

cd <repo-root>/mcp-server
npm install
npm run build

Requires Node.js 20+. If the video_* tools were unavailable in this session, the host must restart its session after the build (and, on Codex manual setups, after the config.toml wiring in "Host-specific final steps") for the MCP server to connect — tell the user this explicitly, and continue setup in the new session.

Step 1: Engine Selection

Ask the user:

Which local transcription engine do you want?

a) parakeet (recommended) — NVIDIA Parakeet TDT 0.6B v3, ONNX int8 on CPU. ~5x realtime, English + 24 European languages. Needs a one-time ~670MB model download and a small Python venv.

b) whisper-cpp — whisper.cpp (whisper-cli). ~1–2x realtime, 99 languages. Models auto-download on first use.

c) whisper-python — openai-whisper CLI. Slowest, 99 languages. Use only if you already have it installed.

All engines use ffmpeg to extract video frames — your agent sees the frames directly.

Step 2: Dependency Check

Call video_setup with the chosen engine. Report exactly what is present and what is missing (ffmpeg, whisper-cli, Python venv, Parakeet model dir, yt-dlp). Mention that yt-dlp is optional and only needed for YouTube URLs.

If the video_* tools are not available at all, jump to "Host-specific final steps" below to wire up the MCP server first, then come back here.

For anything missing, walk through the relevant section below.

Step 3a: Parakeet Setup (engine = parakeet)

First ask:

Do you already have a Parakeet ONNX model directory on disk (for example from the Handy app)?

a) Yes — give me the path

b) No — download it for me (~670MB)

If yes: call video_configure with parakeet_model_dir set to their path. Skip the model download below, but still verify the venv.

If no (or venv missing), guide through these exact commands.

1. Create the venv:

python3 -m venv ~/.agent-video-vision/venv

If that fails (missing python3-venv / "ensurepip is not available"), use uv instead:

uv venv ~/.agent-video-vision/venv

2. Install onnx-asr (note the onnxruntime pin — newer versions break onnx-asr):

~/.agent-video-vision/venv/bin/pip install "onnx-asr[cpu,hub]" "onnxruntime<1.24"

uv variant:

uv pip install --python ~/.agent-video-vision/venv/bin/python "onnx-asr[cpu,hub]" "onnxruntime<1.24"

3. Download the int8 model from HuggingFace (istupakov/parakeet-tdt-0.6b-v3-onnx):

~/.agent-video-vision/venv/bin/python -c "
from huggingface_hub import snapshot_download
snapshot_download(
  'istupakov/parakeet-tdt-0.6b-v3-onnx',
  allow_patterns=['*.int8.onnx', 'vocab.txt', 'config.json'],
  local_dir='~/.agent-video-vision/models/parakeet-tdt-0.6b-v3-int8'.replace('~', __import__('os').path.expanduser('~')),
)
print('done')
"

Offer to run these commands for the user. Run them one at a time and confirm each succeeds before the next.

WSL note: if the user points at a model dir under /mnt/* (a Windows drive), warn that it loads much slower than the Linux filesystem and suggest copying it to ~/.agent-video-vision/models/.

Step 3b: Whisper Setup (engine = whisper-cpp or whisper-python)

• whisper-cpp: needs the whisper-cli binary (brew install whisper-cpp on macOS, or build whisper.cpp from source on Linux). Models auto-download from HuggingFace (ggerganov/whisper.cpp) on first use — no manual download.
• whisper-python: needs the whisper CLI (pip install openai-whisper).

Then ask:

Which whisper model?

a) tiny (75MB) — very fast, basic quality

b) small (500MB) — good balance

c) large-v3-turbo (1.5GB) — best cost-benefit, 8GB+ RAM

d) large-v3 (2.9GB) — maximum quality, 16GB+ RAM

e) auto — pick based on your hardware

Call video_configure with whisper_model.

Step 4: Frame Configuration

Ask these one at a time (skip if user says "defaults"):

Resolution

Frame extraction resolution (width in pixels, height auto-scales)?

a) 256px — low res, fast, fewer tokens

b) 512px (default) — good balance

c) 768px — higher detail

d) 1024px — maximum detail, more tokens

Call video_configure with frame_resolution.

Frame Mode

How should frames be delivered?

a) Images (Claude Code default) — frames come back as inline images in the tool result (better perception, more tokens)

b) Paths (Codex default) — the tool returns timestamp -> /abs/path file lists; the agent opens the frame files it needs

c) Descriptions — frames are converted to text descriptions by the host agent (fewer tokens, loses visual nuance)

Call video_configure with frame_mode. The host-specific section below has the right default per host.

Step 5: Save Engine

Call video_configure with the chosen engine (plus parakeet_model_dir if not already saved). Confirm the resulting config back to the user.

Step 6: Host-specific final steps

Do the section matching the host you are running on.

If you are Codex

1. MCP server connection. If the video_* tools are not available in this session, the MCP server isn't connected. Add this block to ~/.codex/config.toml (use the absolute path to the plugin/clone directory):

   [mcp_servers.agent-video-vision]
   command = "node"
   args = ["/home/YOUR_USER/agent-video-vision/mcp-server/dist/index.js"]  # absolute path — ~ does not expand in TOML
   

   Then tell the user to restart the Codex session so the tools load.

2. Install the frame-describer custom agent. Codex plugins cannot bundle agents, so copy the TOML into the personal agents dir:

   mkdir -p ~/.codex/agents
   cp ~/agent-video-vision/codex/agents/frame-describer.toml ~/.codex/agents/
   

3. Recommend paths mode. Codex cannot see inline MCP images — it must open frame files with its view_image tool. Save it:

   video_configure { "frame_mode": "paths" }
   

If you are Claude Code

Nothing extra — the plugin already provides the MCP server, skills, and the frame-describer subagent. Keep frame_mode: "images" (the default) so frames come back inline.

Step 7: Verify

Run video_setup once more — everything should now report OK.

Then ask:

Setup complete! Want to run a quick smoke test? Give me a path to any short video file.

If the user provides one, call video_info on it, then video_analyze with transcription: true to confirm the engine actually transcribes. Report duration, resolution, and the first transcription line.

Important

• Ask ONE question at a time — never combine multiple questions
• Call video_configure after EACH answer to save incrementally
• If the user says "default" or "just use defaults", set engine to parakeet with defaults, jump to Step 2, then still do Step 6 for your host
• If the user seems experienced, keep it concise — don't over-explain