Openclaw Replicat skill

SKILL.md

name	replicate
description	Search, explore, and run ML models on Replicate (image gen, video, audio, text, etc.)
homepage	https://replicate.com
metadata	openclaw requires primaryEnv bins env curl jq REPLICATE_API_TOKEN REPLICATE_API_TOKEN

Replicate

Run state-of-the-art open-source and proprietary ML models via the Replicate cloud API.

API key

• REPLICATE_API_TOKEN env var
• Or set skills.entries.replicate.apiKey / skills.entries.replicate.env.REPLICATE_API_TOKEN in ~/.openclaw/openclaw.json

Key concepts

• Models are identified as owner/name (e.g. black-forest-labs/flux-schnell).
• Official models are always warm, predictably priced, and have stable APIs. Use the owner/name format.
• Community models require a version ID: owner/name:version_id.
• Predictions are individual runs of a model. States: starting -> processing -> succeeded/failed/canceled.

Operations

Search for models

curl -s -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/search?query=QUERY" | jq .

The response includes .models[] with model.owner, model.name, model.description, model.run_count, and .metadata.tags. Also check .collections[] for related curated lists.

To enrich results with official status and creation date:

curl -s -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/models/OWNER/NAME" | jq '{is_official, created_at}'

Get model schema (always do this before running a model)

curl -s -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/models/OWNER/NAME" | jq '{
    description,
    latest_version: .latest_version.id,
    input_schema: .latest_version.openapi_schema.components.schemas.Input.properties,
    output_schema: .latest_version.openapi_schema.components.schemas.Output
  }'

Run a model

For official models (owner/name, no version):

curl -s -X POST -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait=60" \
  "https://api.replicate.com/v1/models/OWNER/NAME/predictions" \
  -d '{"input": {INPUT_PARAMS}}' | jq .

For community models (owner/name:version_id):

curl -s -X POST -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait=60" \
  "https://api.replicate.com/v1/predictions" \
  -d '{"version": "VERSION_ID", "input": {INPUT_PARAMS}}' | jq .

The Prefer: wait=60 header makes the API wait up to 60 seconds for the prediction to complete before returning. If the prediction is still running, the response will have status: "processing" and you need to poll.

Poll for results

If status is not succeeded or failed, poll every 2-3 seconds:

curl -s -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/predictions/PREDICTION_ID" | jq '{status, output, error, logs, metrics}'

Keep polling until status is succeeded, failed, or canceled. Print the .logs field on each poll to show progress.

Cancel a prediction

curl -s -X POST -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/predictions/PREDICTION_ID/cancel"

Check account

curl -s -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/account" | jq '{username, name, type}'

Workflow

When a user asks you to run a model or perform a task:

1. If they haven't specified a model, search for appropriate ones
2. Get the schema to understand the model's inputs before running it
3. Run the model with the correct inputs -- use Prefer: wait=60 for sync, then poll if needed
4. Present the output clearly -- for image URLs, display them; for JSON output, format it nicely
5. Link to the prediction: https://replicate.com/p/PREDICTION_ID

Understanding user requests

Users often ask to generate things without specifying how. Plan which models to use.

Examples

• "Make a commercial": Generate keyframes with an image model (FLUX, Nano Banana Pro), then generate video segments between keyframes in parallel with a video model, then stitch them together.
• "Decorate my room": Ask for a room photo and style preferences, then use an image editing model with the inputs.
• "Celebrity selfie video": Get a user selfie, find celebrity images (via web search if needed), generate keyframes, create video segments, stitch together.
• "Summarize a PDF": Use an OCR model on Replicate to extract text, then summarize it yourself (no need for a language model on Replicate).
• "Translate a video": Extract audio -> speech recognition -> translate text yourself -> text-to-speech -> lip sync model with original video.

Finding the right model

• Always use the latest version of a model (Kling 3.0 over Kling 2.5, Nano Banana Pro over Nano Banana, etc.)
• Prefer recently released models over old ones
• Prefer official models over community models
• High run counts can indicate old popular models, not necessarily the best current option
• The search API does its best to surface recent state-of-the-art models, but apply judgment

Downloading outputs

When a prediction returns file URLs (images, audio, video), download them locally:

curl -sL "OUTPUT_URL" -o output_filename.png

Report the saved path so the user can access the file. Use descriptive filenames with timestamps when generating multiple outputs.