RunPod Media Skill

Generate AI images and videos using RunPod public endpoints. All output is saved to ~/runpod-media/.

API Keys

One key required — add to ~/.openclaw/secrets.json:

Key path	Purpose	Get it from
INLINECODE2	Call RunPod endpoints	runpod.io/console/user/settings

Local images are uploaded to Cloudflare R2 as presigned URLs (1 min expiry) before being sent to RunPod endpoints. R2 credentials are read from /cloudflare/r2 in secrets.json — already configured ✅

imgbb is no longer used. R2 presigned URLs replace it for all local file uploads.

R2 cleanup: Objects in uploads/ are auto-deleted after 1 day via a lifecycle rule on the openclaw bucket. Presigned URLs expire after 1 min (no access), objects are cleaned up within 24h.

Keys are resolved in this order:

1. 1. OpenClaw secrets.json — ~/.openclaw/secrets.json ✅ (already configured)
2. Env vars — INLINECODE7

How Users Ask (Natural Language Examples)

The user will never type CLI commands — translate their natural requests into the right script call.

Generate an image:

• - "Generate an image of a samurai cat in neon Tokyo" → INLINECODE8
• "Make me a 16:9 image of a stormy ocean at sunset" → INLINECODE9
• "Create an image using Nano Banana — a futuristic city" → INLINECODE10

Edit an image:

• - "Edit this image — add snow falling" → INLINECODE11
• "Use Qwen to edit this photo, make it look like a painting" → INLINECODE12

Animate to video:

• - "Animate this image — slow camera pan" → INLINECODE13
• "Make a video from this with Kling" → INLINECODE14
• "Turn this into a 10 second clip with Sora 2" → INLINECODE15

Text to video:

• - "Generate a video of a wolf howling at the moon" → INLINECODE16

List available models:

• - "What image/video models do you have?"
• "List the available endpoints"
• "Show me what RunPod models are available"

→ Run list_endpoints and summarize the output in plain language for the user

Add a new endpoint:

• - "Add this RunPod endpoint: https://console.runpod.io/hub/playground/voice/kokoro-tts"
• "Probe and add these endpoints: kokoro-tts, flux-kontext-pro"

→ Run discover_endpoints add --candidates "<url-or-id>"

---

Capabilities & Cost

Task	Command	Cost	Time
Text → Image	INLINECODE19	~$0.005/image	3–8s
Edit image(s)

edit_image | ~$0.005/image | 5–15s |
| Image → Video | image_to_video | $0.03–$0.90/clip | 30–120s |
| Text → Video | text_to_video | $0.04–$1.22/clip | 30–120s |
| Any endpoint | call_endpoint | varies | varies |

The built-in commands use default endpoints. For more models (Nano Banana Pro, FLUX, Sora 2, Kling, TTS, etc.) use call_endpoint with any RunPod public endpoint ID.

Endpoint Registry

All known public endpoints are in scripts/endpoints.json. List them:

CODEBLOCK0

Call Any Endpoint

CODEBLOCK1

Examples:

CODEBLOCK2

Adding New Endpoints

When the user asks to use an endpoint not in the registry, or the runpod skill reveals a new one:

1. 1. Call it directly with --endpoint <id> — no registry entry needed
2. Optionally add it to scripts/endpoints.json for future sessions

With runpod skill: Use the runpod skill to browse/discover endpoint IDs on the RunPod hub, then pass that ID to call_endpoint here.

Generate Image

CODEBLOCK3

Edit Image

CODEBLOCK4

• - Accepts 1–5 images (local paths or URLs)
• Local files are auto-uploaded via imgbb (requires /imgbb/apiKey in secrets.json)

Animate Image → Video

CODEBLOCK5

Models:

• - wan25 (default) — WAN 2.5, ~$0.026/5s
• INLINECODE34 — Kling v2.1 Pro, $0.45/5s (highest quality)
• INLINECODE35 — Seedance 1.0 Pro, ~$0.12/5s

Generate Video from Text

CODEBLOCK6

Models:

• - wan26 (default) — WAN 2.6, ~$0.04/5s
• INLINECODE37 — Seedance 1.0 Pro, ~$0.12/5s

Defaults

• - Delete after send — always delete the local file after successful delivery. Only keep if the user explicitly asks ("keep it", "save it", "--keep").
• Captions — keep them short and natural. Do NOT include render time or cost unless the user asks. Example: 🦊 Fox under the aurora not 🦊 Fox — 105s render (~$0.026).

Delivering Media to the User

After generating an image or video, always deliver it to the user via their active channel.

The Problem

The message tool with a local media path may fail in sandboxed agent modes due to SecretRef resolution not being available for media sends. This is a known OpenClaw limitation.

The Solution: Use curl + Telegram Bot API directly

Read the bot token from secrets and send via curl — this always works regardless of sandbox mode:

CODEBLOCK7

Where to find CHAT_ID: Use the chat_id from the inbound message metadata (e.g. telegram:1231438498 → use 1231438498).

Delivery Checklist

1. 1. ✅ Generate the file (note the saved path from script output)
2. ✅ Send via message tool with a short, natural caption (no cost/time unless asked)
3. ✅ Delete the local file after successful send: INLINECODE46
4. ⏭️ Skip deletion only if user said "keep it" / "save it" / passed INLINECODE47

Why not message tool?

Try it first — if it works, great. If it returns a SecretRef error, fall back to the curl method above.

Notes

• - Video generation takes 30–120 seconds — inform the user to wait
• Output files saved to ~/.openclaw/workspace/runpod-media/ — accessible in both sandboxed and elevated agent modes
• Shared utilities in scripts/_utils.py — do not call directly