Frequently asked questions
Everything about Overshoot's real-time AI vision infrastructure, straight from the documentation. Last updated July 15, 2026.
What is Overshoot?
Overshoot is real-time AI vision infrastructure. It runs Vision Language Models on live video: point it at a video source, describe what you want in plain English, and get results continuously as the video plays, as fast as 200 ms. Overshoot is backed by Y Combinator.
How does Overshoot work?
Create a stream, publish your video into it over LiveKit, then ask questions about it with an OpenAI-compatible chat completions call. You describe what you want in plain English and reference the frames or segments you care about, and results stream back as the video plays. The prompt is the program: there are no models to train and no vision pipelines to build.
Which video sources does Overshoot support?
You publish frames to a stream over LiveKit (WebRTC), so any
LiveKit publisher works: a browser webcam or screen share, a
native app, or a server-to-server feed. In the browser you
connect with the
livekit-client
library using the room URL and token returned when you create
the stream.
How fast is Overshoot?
Results stream continuously as the video plays, with latency as low as 200 ms.
How do I integrate Overshoot?
Overshoot is a REST API at
https://api.overshoot.ai/v1beta.
There is no Overshoot SDK to install: you create a stream,
publish frames over LiveKit, and query them through an
OpenAI-compatible
/chat/completions
endpoint, referencing frames and segments with
ovs://
stream URLs. Any HTTP client and any LiveKit publisher will
do. See the
API reference.
Which AI models can I use with Overshoot?
Overshoot runs a range of Vision Language Models.
Overshoot-hosted models are the fast path, with IDs like
Qwen/Qwen3.6-27B-FP8
and
google/gemma-4-31B-it.
Proprietary models from Google, Anthropic, and OpenAI are
available as a passthrough. List everything with an
unauthenticated
GET /v1beta/models
call, use the models marked
ready,
and switch models by changing the
model
parameter in your request.
How much does Overshoot cost?
Overshoot uses prepaid credits and bills per token, with
separate input, output, and cached-input rates that vary by
model. Pricing is public and unauthenticated at the
/billing/pricing
endpoint. Cost scales with how many frames you send and their
resolution, so a 480p stream and single-frame questions cost
far less than full-resolution video segments.
What can I build with Overshoot?
OCR on camera feeds; real-time scene description for blind and low-vision users; security and monitoring on live camera feeds; sports and fitness analysis; robotics and computer-use agents that see and act on screen or camera input; retail analytics; and structured JSON extraction from video using an output schema.
How do I get started with Overshoot?
Get an API key at platform.overshoot.ai/api-keys, create a stream, publish your camera over LiveKit, and send your first chat completion following the quickstart at docs.overshoot.ai. You can also join the waitlist at overshoot.ai or contact founders@overshoot.ai.