The open runtime, hosted.

Percept Cloud runs the same OSS engine you can install locally. Successful run results come only from the installed provenance-stamped percept-vision wheel — results are not mocked or predetermined, and zero incidents is a valid result.

How a hosted run works

1. AuthenticateEvery request except the health check carries a bearer token — the meeting demo token or a provisioned client API key.
2. Pick mediaUpload your own video (MP4/MOV/WebM/MKV) or audio (WAV/MP3) — limited to 10 minutes (600 s) — or use a hosted sample from GET /v1/samples — currently red-panel, with a suggested goal.
3. State a goalA plain-language condition, 3–500 characters — the same verbatim contract as the OSS watch() seam.
4. Create and pollPOST /v1/runs returns 202 with a poll URL; the run moves queued → running → succeeded | failed.
5. Inspect the resultWhile the run is running, each poll streams a partial trace live so you watch the ledger fill instead of waiting on a blank panel. A succeeded run carries incidents, the final trace, and an OssProvenance block stamped result_kind: real_oss_execution.

Client API keys

Provisioned client API keys are usable as bearer tokens alongside the meeting demo token — same scheme, same endpoints: Authorization: Bearer <token>. To get a key for your team, write to keys@velvee.ai.

Provider selection

Vision judging runs on a selectable provider per run: vision_provider=gemini (default) or vision_provider=openrouter. This mirrors the OSS posture that providers are identity and models are configuration — the cognition semantics do not change with the vendor.

Source profile and execution policy

source_profile tells the runtime what kind of footage it is watching: physical_scene (default) for cameras and the real world, or screen_document for screen captures and documents. Optionally, execution_policy biases how a run spends its budget — balanced, burst, low_cost, or privacy_edge. Both are plain multipart form fields on POST /v1/runs; leave execution_policy unset to accept the service default.

Audio and fused evidence

Standalone audio uploads are accepted: a WAV or MP3 file as media, judged with evidence_modality=audio. The default modality is fused, which judges across vision and audio together; vision restricts judging to frames alone.

An honest quiet run

A succeeded run with incidents: [] is a valid result — the goal's condition never became true, and the system does not fabricate a fire. A quiet run is an answer, not a failure.