VoiceRun Voice Agent Platform
Voice agents, built as software
Most voice AI sends a transcript to an LLM and prays. VoiceRun agents are code: your team, or your coding agent, writes the logic, picks the best speech and language models for each turn, and ships with the recordings, traces, evals and simulations to prove it works.
async def handler(event: Event, context: Context): if isinstance(event, StartEvent): yield TextToSpeechEvent(text="Thanks for calling. How can I help?") if isinstance(event, TextEvent): said = event.data.get("text", "") # The long tail lives in code, not in a prompt if asks_for_a_person(said): yield TextToSpeechEvent(text="Connecting you now.") yield TransferSessionEvent( phone_number=context.variables.get("FRONT_DESK"), ) return async for sentence in answer(context, said): yield TextToSpeechEvent(text=sentence)Orchestrating the world's best speech and language models
Vertex AI
AWS Bedrock
DeepSeek
Telnyx
Why code
We don't prompt and pray
A prompt gets you a demo. Production is the long tail: the caller who changes their mind, switches language, or asks for something nobody planned. Chasing it takes software engineering.
Prompt and pray
VoiceRun
Every edge case is another paragraph in the prompt.
Edge cases are branches you can read, test and fix in one place.
Locked to whichever speech and LLM vendors the platform picked.
Choose models per agent and per turn, with failover when a provider stumbles.
You find out it broke when a customer complains.
Evals score every call, and simulations catch regressions before release.
A change ships to every caller at once.
Immutable releases, environments, and weighted traffic splits for canaries.
Feature highlights
Everything it takes to run voice agents in production
Short tours of the platform and the CLI, from the first line of code to the phone ringing. Each one stands on its own.
Agents are Python, not a prompt in a text box: branch on what the caller said, call your systems, and handle the long tail of real calls in code you can test.
Built for coding agents
Your coding agent already knows VoiceRun
vr setup connects Claude Code, Codex, or your coding agent of choice to the VoiceRun MCP server and installs skills for building, testing, releasing and debugging agents. Everything in the console is a command, so the agent can run the whole loop and read the results.
- Scripted, headless conversations that return JSON
- Simulations and evals from the command line
- Session transcripts, events and traces for root-causing a bad call
$ pip install voicerun-cli$ vr setup # MCP server + vr skills for Claude Code, Codex, or your coding agent$ vr signin✻ Claude CodeBuild a support line that books appointments and transfers to the front desk.$ vr debug --headless --script calls.json # scripted conversation, JSON out$ vr simulate staging --name happy-path --wait$ vr release production --entrypoint support-line --weight 25$ vr session trace <session-id> # what happened, turn by turn
3 days
from API access to the first voice-completed test booking
Multilingual restaurant-reservation line
~87%
of calls resolved without a human, across five markets
Same deployment, 30 days of production calls
1.1s → 150ms
happy-path model latency after fine-tuning on production calls
VoiceRun post-training pipeline
“VoiceRun took us from zero to production deployment in weeks. Then their tooling and AI evaluations show us exactly what to improve each week, so accuracy and customer satisfaction keep climbing.”
What we hear
From teams putting voice agents on the phones
“Our agent demos well and falls apart on real callers.”
Real calls are a long tail. In code, each new case is a branch you add, test and release, not another paragraph a model may or may not follow.
A code-first canvas“Our call flows outgrew the no-code builder.”
VoiceRun agents are Python projects in your repo. Call your own systems, keep state across the call, and hand off to people or other agents.
A code-first canvas“We’re stuck with one speech vendor.”
Pick speech-to-text, turn-taking and voice per agent, switch models mid-call, and fail over automatically when a provider has a bad day.
The best audio models, orchestrated“When a call goes wrong, nobody can tell us why.”
Open the session: the recording, what the caller actually heard, every event, and a trace of each model call with its latency.
See everything that happened“Every prompt change feels like a gamble.”
Run simulated callers against the new release, let evaluators grade every conversation, then send it a slice of live traffic before everyone.
Simulate before you ship“It’s too slow, and the big models are expensive at our volume.”
Post-train a smaller open model on your own calls. In one production deployment, happy-path latency fell from about 1.1 seconds to 150 milliseconds.
Fine-tune for the last mile“We can’t rip out our phone system.”
You don’t have to. Bring your Twilio or Telnyx account or your contact center, keep your numbers, or buy new ones from us.
Telephony your way“Our engineers already build with Claude Code.”
One command sets up the VoiceRun MCP server and skills for Claude Code, Codex, or your coding agent of choice, so it can build and debug voice agents too.
Built for Claude Code and Codex
FAQ
Questions engineering leaders ask us
Who owns the agent?
You do. An agent is a Python project with a handler and declarative config that lives in your repository. Your code remains your code, your data remains your data, and the models we train for you serve only you.
Do we have to change carriers or phone numbers?
No. Connect your own Twilio, Telnyx or Infobip account, route from your contact center, or buy numbers through VoiceRun. Phone, web and in-app entrypoints all point at the same releases.
Which models can we use?
Speech-to-text from Deepgram, OpenAI, ElevenLabs, Cartesia, Soniox, Qwen and more; voices from 13 TTS providers plus custom cloned voices; and LLMs from OpenAI, Anthropic, Google, AWS Bedrock and Alibaba through one completions proxy with fallbacks. You can mix them per agent and change them mid-call.
Can we keep our own agent logic or LLM stack?
Yes. In relay mode VoiceRun handles telephony, speech-to-text and voice while your own WebSocket service does the thinking, with a failover URL if it goes down.
How do you handle sensitive call data?
PII redaction can be switched on per deployment for transcripts, events and traces. Retention is configurable from 30 days to 10 years, recordings are off unless you turn them on, and you choose the region an agent runs in.
Do we need our own voice AI engineers?
It helps to have someone who writes Python, and coding agents do much of the work. Our forward-deployed engineers can also build alongside your team, from the first prototype through production.
Put your first agent on the phones
Tell us about the calls you want to automate. We'll scope the agent with you, and our engineers can build it alongside your team.
Your code stays yours · Your data stays yours· Models we train for you serve only you