30-second failover demo

Three scripts, two terminals, one screen recording.

Setup (~2 minutes)

cd OrcaRouter-Lite
cp .env.example .env

# Configure BOTH providers so failover has somewhere to go.
echo 'OPENAI_API_KEY=sk-...'        >> .env
echo 'ANTHROPIC_API_KEY=sk-ant-...' >> .env
# Optional: GOOGLE_API_KEY, GROQ_API_KEY for more failover variety

docker compose up -d
docker compose logs api 2>&1 | grep 'sk-orca-'
# → copy the printed sk-orca-* key

Recording (~30 seconds)

Open three tiles side-by-side for the camera:

1. Terminal A (top) — runs scripts/demo.sh in a tight loop
2. Browser (right) — http://localhost:8000/, Analytics tab visible
3. Terminal B (bottom) — fires scripts/demo_kill_openai.sh mid-recording

# Terminal A — start the loop
export ORCA_API_KEY="sk-orca-..."
./scripts/demo.sh

You'll see output like:

[  1] MISS         gpt-4o-mini                              ok
[  2] HIT          gpt-4o-mini                              ok
[  3] HIT          gpt-4o-mini                              ok
...

After ~10 seconds, in Terminal B:

./scripts/demo_kill_openai.sh

The next request in Terminal A retries via Anthropic; subsequent requests show the resolved model flipping (claude-3-5-haiku-latest etc) without any code change in Terminal A:

[ 11] MISS         claude-3-5-haiku-latest                  ok
[ 12] HIT          claude-3-5-haiku-latest                  ok

The dashboard's recent-requests table fills with provider=anthropic rows; the savings tile updates live.

Capture

• macOS: Cmd-Shift-5, then ffmpeg -i input.mov -vf "fps=12,scale=1200:-1" demo.gif
• Linux: peek or byzanz-record
• Cross-platform: OBS (overkill for 30s but bulletproof)

Trim to ≤30s. Export under 8 MB so GitHub renders inline in the README.

Storyboard

Time	What viewers see
0-3s	Title card: "OrcaRouter Lite — failover in real time"
3-12s	Tight loop succeeding via gpt-4o-mini. Dashboard fills.
12-15s	Terminal B runs demo_kill_openai.sh — visible failover trigger
15-25s	Same loop continues; resolved model flips to claude-* mid-stream
25-30s	Closing card: "No code change. model="auto" did the routing."

Restore

./scripts/demo_restore_openai.sh

Where to publish

• README hero (replace the placeholder image link)
• Show HN cover image
• Twitter/Threads launch post
• ProductHunt gallery

Why this lands

Most LLM router pitches show a config table. This shows a failure happening in real time and being absorbed transparently. The viewer's brain resolves "so this is what would have prevented my 3am pager last month" and that's the share-worthy moment.