The real-time video agent engine
Build AI that sees what’s happening and knows when to help.
A video-call engine. Your models, your expertise, your app.
Video inputA cream blazer.
Dark denim. Clean proportions.
“The blazer works. Keep the darker denim—it gives the look structure.”
One continuous loop
Perception is a loop, not a prompt.
Video, voice and reasoning, working together.
So your agent can stay with the task as it changes.
Video frames → image preparation → current visual context
camera.capture_fps8Reported prototype benchmarks
9.8×
faster responsesvs GPT-6 Luna
8×
higher frame processing ratevs Gemini Live Video
>2.5×
cheaper per minutevs GPT-Live Voice
Figures from our internal prototype, reported in the Second Pair pitch deck. Workload and model dependent; not independent benchmarks or performance guarantees.
One engine. Any specialty.
Cook, Handy and Coach are not separate products. Each is a configuration: a prompt, a set of tools and a proactivity policy on the same engine. Yours can be anything a camera can help with.
Another pair of eyes in the kitchen.
The configuration that makes it
lookup_ingredientcurrent_recipe_stepyour application’s own functionsBuilt for developers
Choose separate speech, vision and reasoning models, set reasoning effort, decide how many frames are captured and analysed, register your own tools and set how proactive the agent may be. Run it through the hosted API or inside your own service.
from secondpair import (
AgentConfig, CameraConfig, ModelConfig, Models, PerceptionConfig, ProactivityConfig,
)
from secondpair.client import SecondPairClient
config = AgentConfig(
name="Cook AI",
prompt="Help the user cook. Ask what they are making; give concrete next steps.",
models=Models(
speech=ModelConfig(name="openai:gpt-realtime"),
perception=ModelConfig(name="provider:vision-model"),
reasoning=ModelConfig(name="provider:reasoning-model", reasoning="high"),
),
camera=CameraConfig(capture_fps=8, processing_fps=4),
perception=PerceptionConfig(enabled=True, analysis_fps=0.5),
proactivity=ProactivityConfig(enabled=True, min_confidence=0.85),
)
# Inside your authenticated backend route. Keys stay in server secrets.
async with SecondPairClient(base_url=SECONDPAIR_API_URL, api_key=SECONDPAIR_API_KEY) as sdk:
credentials = await sdk.create_session(user_id=user.id, config=config)SDK v0.1 — Python engine, TypeScript client and React binding. Packages are shared with early-access teams; they are not on a public registry, so don’t install similarly named packages. Model names above are placeholders for deployments you have access to.
Build with Second Pair
We’re onboarding a small number of teams building live, camera-first experiences. Tell us what you want to build; we’ll reply personally.