🍳Recipes

Practical scripts built from the commands in the Command Reference. Every example uses --json where the output needs to be parsed by another command.

Batch generate for social media#

Different aspect ratios for different platforms, run back to back:

picx image "minimal flat-lay coffee, aesthetic" --aspect-ratio 1:1       # IG post
picx image "bold dramatic thumbnail" --aspect-ratio 16:9 --size 2K       # YouTube thumbnail
picx image "behind the scenes shot" --aspect-ratio 9:16                  # IG story

Batch edit a folder#

Upload every image in a directory, then apply the same edit to each:

for f in ./photos/*.jpg; do
  url=$(picx upload "$f" --json | jq -r .url)
  picx image edit "vintage film grain" -i "$url"
done

-i also accepts a local path directly and uploads it for you, so the loop above can drop the explicit picx upload step:

for f in ./photos/*.jpg; do
  picx image edit "vintage film grain" -i "$f"
done

This requires jq to parse the --json output. If you don't have it installed, parse the output with your shell's own JSON tooling instead — the shape is stable and documented in the Command Reference.

Poll a video generation to completion#

job_id=$(picx video "a drone shot flying over a coastline" --json | jq -r .id)

while true; do
  status=$(picx job "$job_id" --json | jq -r .status)
  echo "status: $status"
  [ "$status" = "completed" ] && break
  [ "$status" = "failed" ] && { echo "generation failed"; exit 1; }
  sleep 10
done

picx job "$job_id" --json | jq -r .output_url

Lip-sync a talking-head clip to a voice track#

lipsync mode takes no prompt — the audio drives the mouth. It needs both a source video and an audio track:

clip=$(picx upload ./presenter.mp4 --json | jq -r .url)
voice=$(picx upload ./voiceover.mp3 --json | jq -r .url)

job_id=$(picx video --mode lipsync \
  --source-video "$clip" --audio "$voice" --json | jq -r .id)

picx job "$job_id" --watch

Replay a failed webhook delivery#

When a delivery to your endpoint failed (endpoint was down, returned a 5xx), find it and re-fire the same signed payload. The redelivery is a real outbound POST — only run it when you're ready to receive it:

# find the failed delivery for a webhook endpoint
del_id=$(picx webhook deliveries wh_abc123 --json \
  | jq -r '.deliveries[] | select(.status=="failed") | .id' | head -1)

# re-fire it
picx webhook redeliver "$del_id"

Check credit balance before a large batch#

balance=$(picx balance --json | jq -r .credits)
echo "current balance: $balance"

if [ "$balance" -lt 500 ]; then
  echo "low balance — top up before running the batch"
  exit 1
fi

Connect an AI agent instead of shelling out#

If the consumer of these commands is itself an AI agent (Claude Code, a custom LLM loop, etc.), it's usually better to connect it directly to the PicX MCP server rather than have it shell out to picx commands — the agent gets structured tool calls instead of parsing CLI stdout:

picx mcp install --client claude
picx mcp doctor

Use CLI scripting (the recipes above) for deterministic pipelines you control; use MCP for an agent that decides what to generate based on conversation.

FAQ#

Can I run these recipes in CI?#

Yes — picx-cli has no interactive prompts once PICX_API_KEY is set, so every command here is CI-safe. Store the key as a CI secret, not in the script.

Do these recipes handle rate limits or retries?#

No — they're minimal examples. For production automation, wrap calls in a retry loop with backoff on non-2xx --json output, and check picx usage periodically to stay ahead of your account's rate limit.

What's the difference between using these recipes and connecting via MCP?#

CLI recipes are for deterministic, script-driven pipelines you write once and run repeatedly (a cron job, a CI step, a batch export). MCP is for an AI agent that decides at runtime what to generate based on a conversation — see MCP Overview.