Menace Voice is a platform for building and running conversational voice agents. You shape the conversation as a workflow, test it from your browser, connect it to live data, and then make it available over a phone number or a website. Each completed conversation produces a run record with the transcript, recording, context, usage, and cost information you need to review what happened.
The recommended path
Start with a Web Call. It exercises the speech, model, voice, workflow, and tool pipeline without requiring a phone number. Connect telephony only after the core conversation and its failure paths are working.
Choose your environment#
Use the hosted Menace Voice app when you want to start immediately. A self-hosted deployment gives you control over the application, database, Redis, and object storage, but you are also responsible for Docker resources, provider credentials, networking, upgrades, and backups. The workflow builder behaves the same way in both environments.
- Hosted: sign in, open the guided agent setup, and configure a managed or bring-your-own model stack.
- Self-hosted: deploy the Menace Voice services first, sign in to the local dashboard, then configure model providers before testing an agent.
- Keep API keys, service keys, telephony secrets, and provider credentials in the platform's credential and configuration screens, never in prompts.
Know the five core pieces#
- Agent or workflow: the versioned graph that defines the conversation. The dashboard says agent; the API and SDK often say workflow.
- Nodes and edges: nodes own prompts or actions, while directed edges describe the conditions that move the conversation forward.
- Model configuration: either realtime speech-to-speech, Menace Voice-managed inference, or separate BYOK LLM, voice, and transcriber services.
- Tools and knowledge: tools read or change live systems; knowledge base documents provide reference material; initial context carries data for one run.
- Run: the durable record of one web, inbound, outbound, or campaign execution, including artifacts and gathered context.
Build in this order#
- Define one job, the caller, a successful outcome, and the things the agent must not claim or do.
- Complete the guided setup: Purpose, Knowledge, Behavior, Voice, Connections, and Launch.
- Run several Web Calls against the draft, including a normal request, missing information, an out-of-scope request, and a tool failure.
- Pass caller-specific data through initial context; attach only the documents and tools needed by each conversational node.
- Add a result path such as a webhook, gathered-context fields, dispositions, or QA analysis so downstream systems can use the outcome.
- Publish the reviewed draft, connect website or telephony traffic, then inspect real runs and iterate from evidence.
Define ready before launch#
A saved workflow is not proof that a caller will have a good conversation. Before launch, write down the expected opening, required questions, acceptable tool outcomes, escalation condition, closing message, and final data you expect to see. Review a real run against that list and keep the published agent unchanged until the draft passes.
Start with the guided walkthrough in Build your first voice agent, then use the production guide when you are ready to connect traffic.