Demo Record

Launch Grackle Demo Recording

This skill sets up and launches the self-recording Grackle demo where a Claude Code agent inside Docker controls a browser with Playwright MCP, narrates with PocketTTS MCP, and records its own screen with ffmpeg.

Prerequisites

• Grackle server must be running (node packages/server/dist/index.js or grackle serve)
• Docker Desktop must be running
• ANTHROPIC_API_KEY must be set (or ~/.claude/.credentials.json mounted)

Steps

1. Build all packages

rush build

2. Build Docker images

# Demo recorder image (Playwright + Xvfb + PulseAudio + ffmpeg + PocketTTS)
# No --secret needed: uses PocketTTS built-in voices (no HF_TOKEN required)
docker build -f docker/Dockerfile.demo-recorder -t grackle-demo-recorder .

# Dev environment image (standard PowerLine)
docker build -f docker/Dockerfile.powerline -t grackle-powerline .

3. Nuke database and restart server

The demo must start with a completely clean database — no leftover projects, tasks, or findings from previous runs. Stale data causes the demo agent to see pre-existing tasks instead of creating them fresh on camera.

# Kill the grackle server (it holds the DB lock)
# Windows (Git Bash): find PID with `wmic process where "name='node.exe'" get processid,commandline | grep server`
# then: taskkill /PID <SERVER_PID> /F
# Linux/Mac: kill $(pgrep -f "node.*server/dist/index.js")

# Delete the database
rm -f ~/.grackle/grackle.db ~/.grackle/grackle.db-wal ~/.grackle/grackle.db-shm

# Restart the server
node packages/server/dist/index.js --port=7434 --web-port=3000 &

Also stop and remove any existing containers:

docker stop grackle-demo-recorder grackle-dev grackle-dev2 2>/dev/null
docker rm grackle-demo-recorder grackle-dev grackle-dev2 2>/dev/null

4. Add and provision environments

Since the DB was nuked, environments must be re-added and provisioned:

# Demo recorder with the full AV stack
grackle env add demo-recorder --docker --image grackle-demo-recorder --runtime claude-code

# Dev environments with repo clone (the demo agent creates tasks on these live)
grackle env add dev --docker --image grackle-powerline --runtime claude-code --repo nick-pape/grackle
grackle env add dev2 --docker --image grackle-powerline --runtime claude-code --repo nick-pape/grackle

# Provision all three
grackle env provision demo-recorder
grackle env provision dev
grackle env provision dev2

Verify all three show "connected":

grackle env list

5. Create project and recording task

The demo agent creates its own project and dev tasks live on camera (Scenes 3-4). Only pre-create the "Grackle Demo" project and the recording task.

grackle project create "Grackle Demo"
# Note the project ID from output

Create the demo recording task using the scene script:

DEMO_DESC=$(cat tools/demo-recorder/DEMO-TASK-PROMPT.md)
grackle task create <PROJECT_ID> "Record Demo Video" --env demo-recorder --desc "$DEMO_DESC"

6. Start the recording task

grackle task start <RECORDING_TASK_ID> --model haiku

Important: Use --model haiku to keep the demo fast and cheap.

7. Monitor and extract

After starting the task, monitor the session and automatically extract the video when it completes.

Poll the task status until it leaves in_progress. Then copy the MP4 out:

# Poll task status every 30s until done
grackle task list <PROJECT_ID>
# When task status is "review" or "done":
docker cp grackle-demo-recorder:/workspace/grackle-demo.mp4 ./grackle-demo-podcast-vN.mp4

Also check ffmpeg recording status if needed:

docker exec grackle-demo-recorder bash -c "ls -lh /workspace/grackle-demo.mp4 2>&1"

Troubleshooting

• Chrome deadlock: Never call mcp__playwright__browser_install — Chrome is pre-installed
• No video: Check docker exec grackle-demo-recorder bash -c "cat /tmp/ffmpeg.log"
• Windows line endings: If scripts fail with /bin/bash\r, run sed -i 's/\r$//' tools/demo-recorder/*.sh
• Auth errors after container recreate: Remove and re-provision containers (tokens are invalidated)
• Dev agent model: Always use --model haiku when starting tasks to avoid defaulting to expensive models

Source: nick-pape/grackle — distributed by TomeVault.