Dino CLI builds AI voice agents that make real phone calls. Design the call, test it on a real phone, then run it across thousands of customers.
You lay out the steps of the call once. Dino CLI holds that structure for the whole call, which frees the model to do what it is genuinely brilliant at.
dino agent validate runs the same checks our engine runs before it will accept your agent for a live call. If it passes on your machine, it will load in production.
The same published agent takes one preview call, holds an inbound line, and works a hundred-thousand-row campaign. You swap the campaign type and rewrite nothing.
Most voice agents stuff the whole call flow into the prompt and hope it survives twenty minutes of conversation. Dino CLI runs it as real logic instead, which is what lets the model spend its attention on the person rather than on remembering the plan.
dino agent validate checks three things: that every step can actually be reached, that every outcome points at a step that exists, and that every path has a way to end the call.
Miss one and the agent walks a caller into a dead end mid-call. It goes quiet, or loops the same question, while a real customer waits on the line. Validate catches it on your machine in about a second.
dino agent preview-call places a real, guarded call to your approved number with your real placeholder values. No infrastructure to provision, no telephony account to wire up first.
A transcript tells you the agent said the right words. Hearing it tells you whether anyone would stay on the line.
Point the same agent at a batch for outbound campaigns, a drip for a queue your systems feed live, or an inbound line customers call themselves. Import from CSV, then start, pause, resume, and retry from the terminal.
The agent you previewed is the exact agent that runs at volume: same file, same validation, same engine. Nothing gets rebuilt or reconfigured on the way to production.
These models are extraordinary at conversation. Holding that quality across a million real calls takes a layer on top: turn-taking, timing, confirmations, recovery.
We built that layer and ran more than a million production calls through it. Everything we learned went back in, which is why your first live agent takes two or three days instead of two or three weeks.
Today you can run on Grok, Gemini, or GPT, using the live speech-to-speech model from each. You pick the model in agent.toml and change it with one line, without touching your call flow. We add models as they ship.
Request access and we approve a batch of teams every week. There is no waitlist you sit on indefinitely. If you are building something real, tell us what it is and you will usually hear back within the week.
From the terminal: scaffold a call flow, validate it, ring your own phone to hear it, then run it across a contact list. Every command returns JSON, and the Skills we publish teach Claude Code or any coding agent to write, check, and ship those agents alongside you.
No. Install it, build an agent, and ring your own phone the same afternoon. We provide the telephony for test calls, so there is nothing to provision. You only need your own number when you want customers calling you.
Scaffold an agent, validate it, and ring your own phone before your coffee gets cold.