Use Gizmo from your agent

Give your agent a prompt, optional reference images, and an export format. The Gizmo Asset API returns a job to track and an asset to download. These docs also provide agent-readable docs and an OpenAPI specification. The API supports individual assets, Team batches, and Team scene generation. Scene generation requires a compatible Team release. Use the API base URL configured for your Team; Antim dev testing uses https://antim-pi-test--api-antim-dev-scenes.modal.run. Keep uploads, submission, polling, and downloads on the same base URL.

Setup

Select Personal or your Team in Gizmo, then create a key in API Access.
Make the key available to your agent as GIZMO_API_KEY.
Copy the following instructions into your agent.

You have access to the Gizmo Asset API. Base URL: https://api.gizmo.antimlabs.com
Authenticate with the header "Authorization: Bearer $GIZMO_API_KEY".

- Generate an asset: POST /v1/assets {"prompt": "..."}.
  Save a unique Idempotency-Key before submission; reuse it with the same body
  after a lost response. Team generation requires this header.
  Optionally include "reference_image_urls": ["https://..."] (up to 8 images across URL and uploaded-image references).
- Team batches: POST /v1/batches with a unique Idempotency-Key and
  {"items": [{"custom_id": "item-001", "input": {"prompt": "..."}}]}.
  Accepts up to 1,000 inputs. Save batch_id; poll GET /v1/batches/{batch_id}.
  Read GET /v1/batches/{batch_id}/items and follow next_cursor while has_more is true.
  A completed batch may contain failures. Inspect each item's outcome.
  Items queue automatically within the Team processing limit shared by all its batches and individual generations.
  Do not resubmit queued items; the processing limit is separate from HTTP request rate limits.
- Team scenes: POST /v1/scenes with a saved unique Idempotency-Key and
  {"prompt": "A simple wooden workbench against a plain wall", "max_asset_definitions": 1,
   "max_instances": 1, "max_cost_cents": 1300}.
  Add reference_images with url or inline data_url; no separate upload is needed. The $13 cap in this example assumes
  a $5 Team asset price and an $8 scene assembly fee; use your Team's actual prices.
  Save both job_id and scene_id. Poll the job or GET /v1/scenes/{scene_id}.
  Fixed-priced scenes charge for delivered unique canonical assets plus the scene fee;
  repeated instances do not add asset charges. Failed/cancelled scenes cost zero.
  Read job.scene_pricing for the accepted prices and job.settled_spend_cents for the charge.
  Inspect result.status, sceneCommitted, validationPassed, assetDefinitions and counts.
  A partial or candidate delivery may still be charged; do not call it fully validated.
  Download only available result.exports with POST /v1/scenes/{scene_id}/export?format=blend
  (or glb, usd, usdz, source). Save the binary response after checking its HTTP status.
  Scene exports use scene_id, not the asset export endpoint below.
- Upload local references: POST /v1/images with multipart files fields.
  Use returned IDs in reference_images: [{"image_id": "...", "view": "front"}].
  Upload and generate using the same account. Never expose the API key in frontend code.
- Save job_id from the 202 response. Poll GET /v1/jobs/{job_id}?include_result=true
  every 30-60 seconds until job.status is succeeded, failed, or cancelled.
  Personal-key clients can also open GET /v1/jobs/{job_id}/events for SSE progress.
  Team SSE currently returns 403 workspace_route_unavailable; use polling.
  Generation continues if you disconnect; save job_id and reconnect later.
  Do not rely on email or webhooks for API completion.
- On asset success, read result.asset.id. Get details with GET /v1/assets/{asset_id}.
- Export with POST /v1/assets/{asset_id}/export?format=mjcf (no request body).
  Formats: mjcf, usd, usdz, sdf, glb; availability depends on the accepted release. Save the binary response to a local file:
  .usdz for usdz; .zip for bundles; .glb for a raw native preview. Team export?format=glb is a ZIP;
  GET /v1/assets/{asset_id}/files/glb returns the saved raw visual GLB.
  Check HTTP status and Content-Type before saving.
- Browse generated assets with GET /v1/assets?limit=20; follow next_cursor.
- Cancel with POST /v1/jobs/{job_id}/cancel or POST /v1/batches/{batch_id}/cancel.
  Cancellation is a request; poll until workers have stopped.
- Team keys do not currently support SSE, GET /v1/usage or asset deletion.
  Do not retry 403 workspace_route_unavailable. Use polling for progress and Team Billing for the current balance.
- A successful job means an output was published, not that all physics QA passed.
  Inspect physics-qa.json when present; respect failures, skips and input hashes.
- On failure, report job.error. Estimates are advisory; do not assume a fixed duration.
- Full reference: https://docs.gizmo.antimlabs.com/api-access

Task: generate {describe an asset}, wait for completion, export it for {simulator},
and give me the saved file.

Example tasks

Generate a cabinet with movable drawers and save its MJCF export.
Recreate a chair from a public product image and download a USDZ file.
Generate missing equipment assets from a list and save their USD exports to your project.
Reconstruct a workbench scene from a reference photo and download its Blender file.

For complete inputs, response fields, statuses, and error handling, see API overview and Scene generation.

Start here

Generate

Manage results

Operate

Use Gizmo from your agent

Setup

Example tasks

​Setup

​Example tasks

Setup

Example tasks