Agents watch footage of any length, gather evidence over time, and turn it into clips, data, reports, and real-world actions.
Hand Ambient a schema and it returns clean, typed JSON — tasks, subtasks and steps, detections, events, citations — frame-accurate and ready for your training pipeline.
Ambient finds the moments, cuts and reframes to 9:16, and adds captions and a music bed — a share-ready social reel from a long video.
Ambient writes polished deliverables from video — product documentation, reports, recipes, blog posts — pulling in the right screenshots and citing the source.
Ambient's agents follow a video through time—pausing, searching, and looking closer when something matters.
Moves through time, follows what matters, and looks closer before acting.
Point Ambient at the footage you already have and ask for the output you actually need.
Find unsafe actions, missing PPE, and procedure violations with timestamped evidence.
Reconstruct what happened, in what order, and where—with cited moments and annotated frames.
Turn demonstrations into actions, bounding boxes, click-points, and step-by-step trajectories.
Identify successes and failures, score trajectories, and extract the moments that explain each result.
Review product or process footage against a checklist and surface anomalies with evidence.
Search, segment, label, and annotate large video collections for training and evaluation.
Whether it’s a ten-second clip or hours of footage, Ambient moves through it selectively—finding and examining the moments that matter.
Ambient searches across the full timeline, pauses around relevant moments, and examines them in detail—without flattening the video into a transcript.
Ambient doesn’t pull in hours of footage to examine a few seconds. It fetches only the moments it needs—often just 1–2% of the source.
Not a paragraph you have to re-parse — a structured, verifiable output your code can consume directly.
Every claim links to a timestamp or range in the source. Clip-local times are rewritten to the global timeline, so citations land on the right moment.
Pass a schema; get typed data back. The final answer is steered to your shape, with validation and retry when it must be exact.
Bounding boxes and click coordinates on a normalized 0–1000 grid — drawn and visually verified before they reach your answer.
The agent can run bash + ffmpeg to render clips, extract data, or build a report — then hand back a download link.
Reasoning, tool calls and results, per-request tokens and cost — streamed live over SSE. Full observability, not a black box.
analyze_video plugs into Claude Code or any MCP client — streaming progress, a final answer, and a session id for follow-ups.
Run Ambient locally or distribute it across a cluster. Use Studio interactively, run headless batch workloads, or integrate through the SDK and API. Isolate media execution in per-session sandboxes for safer, independently scalable workloads.
Postgres and Redis keep state outside the process, so any worker can continue any session.
Isolate each session’s media workload in its own sandbox while orchestration and model calls remain under your control.
Use Studio, run batch jobs through the SDK, or embed the API directly into your application.
The installer starts Ambient, Studio, Postgres, and Redis in Docker. You’ll need Docker Compose and an API key for an OpenAI-compatible model.
# install Ambient + Studio — safe to re-run curl -fsSL https://ambient-intelligence-hq.github.io/install.sh | bash # when ready, the installer prints and opens Studio # add a video or paste a YouTube link, then ask: > Investigate the incident and share a report with evidence ?
Or let your coding agent set it up. Paste this into Claude Code, Codex, or Cursor:
Set up Ambient, an open-source video engine for AI agents, on this machine. Read the setup guide and follow it: curl -fsSL https://ambient-intelligence-hq.github.io/agents/setup.md Show me your plan before installing anything
Deploying for a team or cluster? Detailed self-hosting guidance is coming soon.
Ambient