videosense-agent

It watches your videos, reasons about what it sees, and answers with the clip and the chart to prove it.

▶ Try it live — videosense.work

_{English · 简体中文}

_{A real session, replayed — the question is typed, the agent streams its tool steps (one self-repair included), then answers with three real clips from the library.  ▶ Play the interactive demo}

The answer — and what it cost

Ask in plain language; get the conclusion first, the clips that prove it, and the numbers behind the claim. Every reply carries its receipts in a quiet footer: steps, tools, seconds — and dollars.

How it answers

No pre-baked pipeline: the model picks every next move and keeps going until it can prove an answer, streaming each step back live. When a library went missing mid-analysis, it rewrote its own code and finished the job — see that run.

Run it in 30 seconds

export GCP_PROJECT="your-gcp-project"
export REPL_USE_MOCK_DB=1
uvicorn api.server:app --port 8000        # then open http://localhost:8000

_{No database, no cost — a built-in sample library. You only need gcloud auth application-default login for Gemini.}

Built to be trusted

Every change ships through a 370-task automated evaluation — deterministic verifiers (no LLM judge), honesty and safety tasks pinned as must-pass, and answers that refuse to bluff: if the library doesn't have what you asked for, VideoSense says so instead of showing the nearest lookalike. Prompt changes are validated by an evolutionary loop with statistical gates before a human ever reviews them.

License & commercial use

VideoSense is source-available under the Elastic License 2.0: you're free to read, run, and modify the code — but you may not offer it (or a derivative) as a hosted or managed service. For commercial licensing, get in touch.