MIT licensed · Open source

Talk to your videos.

Open-source AI for long-form video Q&A. Self-host on your GPU, or plug into any provider.

MIT licensed · Self-hostable · Bring your own models

Every other tool either caps out after a few minutes of video or ships your footage to someone else's cloud. OpenVideoSearch does neither.

demo.mp4 · coming soon

grounded

Answers that point to the moment

Exact timestamp citation on every answer, no exceptions
Agent verifies against raw frames before replying

local

Your videos never leave your machine

Full offline stack on your GPU: VLM, ASR, embedder, reranker
No data upload, no per-query bill. Or swap in any OpenAI-compatible endpoint.

agentic

Watch the agent work

Searches transcript, frames, and speakers in one loop
Every tool call streams live so you see reasoning as it happens
Pinpoints the speaker and exact frame, not just the clip

indexed

Understands the whole video before you ask

Global summary plus hierarchical time-window index at ingestion
Full speaker inventory with role detection
Holds up across 10+ hour recordings without degrading

Self-host in minutes

Pick your path: bring an API key, or run every model locally on your own GPU.

Read the README →

Local GPU no external API, runs fully offline

# configure local model settings in .env, then:
git clone https://github.com/adnane-errazine/OpenVideoSearch.git
cd OpenVideoSearch && cp .env.example .env
docker compose -f docker-compose.yml -f docker-compose.models.yml up

Cloud API any OpenAI-compatible endpoint

# set your API key in .env, then:
git clone https://github.com/adnane-errazine/OpenVideoSearch.git
cd OpenVideoSearch && cp .env.example .env
docker compose up