Visual observations.
One straightforward API.

Add image and sampled-video analysis to your application. Use the same organization-scoped API key for REST and MCP. Signed-in console users run jobs with their session—no pasted key.

Authentication and limits

Create an API key in the console’s Settings. Send Authorization: Bearer av_…. Keys are revealed once, stored hashed, and revocable. Requests are bounded to 4 MB, 12 frames, and a 30-second timeline. Images must be base64 JPEG/PNG, at most 2048 pixels on either axis; the browser demo downsizes uploads to 640 pixels.

curl --fail-with-body https://autovision.dev/api/v1/analyze \
  -H "Authorization: Bearer $AUTOVISION_API_KEY" \
  -H "Content-Type: application/json" \
  --data-binary @analysis.json

{
  "frames": [
    {"image": "<base64-jpeg-first-frame>", "at_ms": 0},
    {"image": "<base64-jpeg-next-frame>", "at_ms": 1000}
  ],
  "classify": true
}

camera adapter → timed JPEG frames → AutoVision
AutoVision → observations + summary + job_id → your event pipeline

Live pipeline events

Send Accept: text/event-stream to the same analyze endpoint. Pipeline events include run.started, step.added, step.started, step.completed, step.failed, run.completed and run.failed. The Decider dynamically adds actual tool calls. Final data contains the analysis and job ID. Default requests still return ordinary JSON. Cancel the request to abort outstanding work where supported; no stream resume or automatic retry is provided.

Homepage uploads use the bounded anonymous live pipeline. Prerecorded clips replay cached reports without inference charges. Only concise decision summaries and observable results are displayed, never hidden model chain-of-thought.

Discover lenses and AI models

Authenticated GET /api/v1/lenses lists implemented, separate-API and planned lenses plus approved model keys and skills. Discovery has no inference charge. Organization settings choose planner, scene, package and activity models from this registry. Planned adapters are not executable. Operators register and verify models; customers cannot submit arbitrary model IDs. MCP exposes the same discovery through list_vision_lenses.

Build autonomous apps

Your app supplies frames → OpenCV measures objects and brightness → Bedrock selects a bounded inspection → actual tool results change its next decision → your app applies its own approval and notification policy. Examples include camera-event triage, pet observation, and visual inventory review. This API does not grant permission for physical or security actions.

Autonomous inspection

Send enhanced:true with classify:true to opt into a bounded Bedrock-planned inspection using supplied frames. Set enhanced false to disable it; omission inherits organization settings. Follow-ups can crop/re-detect with OpenCV 5, enhance low light with CLAHE and re-detect, or request focused scene assessment. Each follow-up consumes an additional image quota unit. Read inspection.status, review_recommended, reason, and trace. Failures retain initial observations and recommend review. No notifications or external actions are performed. inspect_scene exposes the same loop through MCP.

Handle the result

A review recommendation can create a task in your app or notify its configured operator. Missing models, timeouts and exhausted budgets recommend review while preserving initial observations. A false flag, or no configured trigger, is not proof of safety or absence. Traces retain actual tool measurements and timings; they are available in the private job detail. No arbitrary URLs, shell commands or model downloads are allowed. Configure policy per organization in Settings. Low-light enhancement does not recover details the camera never captured.

Privacy and retention

The anonymous demo does not persist media or job history. Authenticated analyses retain results, actor attribution, and small sampled previews for seven days, with expiry filtering and DynamoDB TTL where configured. Administrators can clear saved jobs and previews from Settings without disconnecting devices or removing API keys. Original clips are decoded and sampled in the browser, not uploaded. Enabling classification sends sampled frames to Bedrock; follow the provider’s data terms.

Errors