Replicate

Purpose

Run AI models on Replicate: look up a model to get its version ID and input schema, create an async prediction, poll its status, and cancel runs. Use when the user wants to run a specific model by owner/name.

Tooling

All commands go through bin/replicate.py:

bin/replicate.py auth                                                  # verify the API key
bin/replicate.py model --name stability-ai/sdxl                        # get version ID and input schema
bin/replicate.py predict --version <version-id> --input-json '{"prompt": "..."}'  # start a prediction
bin/replicate.py status --id <prediction-id>                           # poll prediction status
bin/replicate.py cancel --id <prediction-id>                           # cancel a running prediction

Always call model first: copy the latest_version_id and match your --input-json to its input_schema before calling predict.

Auth

• Provider id: replicate (credential is collected as custom.replicate)
• Collection: API token via the secure credential flow (credentials.request_api_access); created in the Replicate account settings
• Allowed hosts: api.replicate.com
• Status check: bin/replicate.py auth (must return "ok": true)

Operating Rules

1. predict and cancel are writes: confirm the model, version, and input with the user before running, unless standing permission exists. Predictions cost money per run, so warn about cost too.
2. Predictions are async: after predict, poll status until the status reads succeeded or failed.
3. Output file URLs in a prediction expire after 1 hour; download anything worth keeping right away.
4. Never exfiltrate the credential: the CLI only ever handles surrogates (see bin/replicate.py). Do not print, log, or transmit the key value.

Files

• SKILL.md
• bin/replicate.py

Maturity

🧪 Draft: written from Replicate's public API docs; not yet live-tested end-to-end.