Runware compatibility

To move a Runware client over, see Migrating from Runware. This page is the reference for what is served.

Parsec serves Runware’s task API at https://platform.parseclab.ai/v1 for the image and video models it hosts. Requests keep Runware’s shape (a JSON array of tasks with taskType and taskUUID), Runware’s model ids keep working, and answers come back as {"data": […], "errors": […]}.

Either name works in model. Prices are on the models page.

SDK

Runware’s SDKs default to a WebSocket, which Parsec serves at wss://platform.parseclab.ai/v1:

import { createClient } from "@runware/sdk";

const runware = await createClient({
  apiKey: process.env.PARSEC_API_KEY,
  wsBaseUrl: "wss://platform.parseclab.ai/v1",
});
await runware.connect();

const [image] = await runware.run({
  model: "google:4@3",
  positivePrompt: "A watercolor lighthouse at dusk",
  width: 1344,
  height: 768,
});
console.log(image.imageURL);

For REST instead, set transport: "rest" and httpBaseUrl: "https://platform.parseclab.ai/v1". Either way the SDK’s async default works as it does on Runware: it starts the task, then polls getResponse until the result is ready.

Over the WebSocket, the first message authenticates ([{"taskType": "authentication", "apiKey": …}]), ping answers pong, and each task answers on its own as soon as it is done. Idle connections close after 120 seconds, as on Runware. Results of a dropped connection are not replayed; async tasks are kept, so getResponse still finds them.

HTTP

curl https://platform.parseclab.ai/v1 \
  -H "Authorization: Bearer $PARSEC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[{
    "taskType": "imageInference",
    "taskUUID": "a770f077-f413-47de-9dac-be0b26a35da6",
    "model": "google:4@3",
    "positivePrompt": "A watercolor lighthouse at dusk",
    "resolution": "1K",
    "includeCost": true
  }]'

The key can also go in a leading {"taskType": "authentication", "apiKey": …} task, as on Runware.

Video

videoInference is always asynchronous. Send the task, then poll with getResponse and the same taskUUID:

curl https://platform.parseclab.ai/v1 \
  -H "Authorization: Bearer $PARSEC_API_KEY" \
  -d '[{"taskType": "videoInference", "taskUUID": "4f5b…", "model": "xai:grok-imagine@video-1.5",
        "positivePrompt": "An otter surfing at sunset", "duration": 6, "resolution": "720p"}]'

curl https://platform.parseclab.ai/v1 \
  -H "Authorization: Bearer $PARSEC_API_KEY" \
  -d '[{"taskType": "getResponse", "taskUUID": "4f5b…"}]'

The poll answers "status": "processing" until the video is ready, then "status": "success" with videoURL. A video is charged once, when a poll first finds it done; failed videos cost nothing.

Webhooks

Set webhookURL on an imageInference or videoInference task and the task runs in the background; each result is POSTed to that URL as it finishes, the same item getResponse would return. A task that fails POSTs its error object ("status": "error", code, taskUUID). Put any token you check in the URL’s query string, as on Runware.

• Answer with a 2xx. Anything else is retried with backoff (250 ms doubling to 8 s) for up to 30 minutes.
• A delivery can arrive more than once: dedupe on taskUUID and imageUUID/videoUUID.
• The URL must be https and resolve to a public address; redirects are not followed.

What is served

Not served yet, and refused with unsupportedParameter rather than ignored: uploadEndpoint, providerSettings, advancedFeatures, speech, inputs.referenceVideos, model auto, and images uploaded to Runware by UUID (send the image itself or a URL instead). Task types other than imageInference, videoInference, getResponse, authentication and ping are not served either.

Errors

Errors use Runware’s shape, one entry per failed task, with the HTTP status of the first failure when no task succeeded:

{"errors": [{"code": "unsupportedModel", "message": "…", "parameter": "model",
             "taskType": "imageInference", "taskUUID": "…", "status": "error"}]}

Decisions

If you call TypeSafe’s decisions API through Runware (/v1/decisions), Parsec serves the same request at POST /v1/decisions with its own model, tod-picker. Leave model out or set it to tod-picker; other model names are refused.