26 pre-built templates
Install a proven starting point for common call flows, then reshape it on the canvas. Templates are starting geometry, not black boxes.
Product
Design voice agents on a canvas you can actually read.
The agent builder is where PeachDesk agents take shape: typed nodes, conditional edges, and per-stage provider sourcing on one visual canvas. It is the design surface of the voice AI platform, built for teams who want to own how their agents think, not just rent a pipeline.
Three ways to start, one canvas to finish on
Describe and generate
Tell the AI Agent Builder what the agent should do in plain language. It generates a starting flow you can inspect and edit node by node.
Blank canvas
Start empty and compose exactly the agent you want: a multi-step conversation flow, or a single-prompt agent when one strong prompt is the right design.
16 core node types, wired your way
Every flow is a graph of typed nodes and conditional edges, so the logic of a call is visible before it ever runs. Deterministic equation transitions route calls on facts, not model judgment: a condition like a budget threshold takes the same branch every time.
When a flow needs more than conversation, the canvas has the parts: a code node for custom logic, call transfer, agent-to-agent handoff, in-call SMS, and keypad navigation for callers who prefer buttons.
Each agent sources its pipeline stages independently. Speech-to-text, the language model, text-to-speech, and realtime stages can each run on PeachDesk-managed providers, on keys you bring from your own providers, or on self-hosted inference in Local mode. Frontier mode, Local mode, mix and match: two model planes, one platform.
Test the agent before a customer ever hears it
Four test surfaces cover the distance between a first draft and a production call.
Text chat
Fast logic checks in the editor. Text chat skips the speech stages, so use it for flow, not for voice.
AI caller personas
Simulated callers stress-test the flow with personalities of their own, so awkward branches surface before real callers find them.
In-browser voice
A live WebRTC call with a real-time transcript, exercising the full speech pipeline without a phone.
Real phone calls
Call the agent through your connected telephony and hear exactly what a customer will hear.
Before any run, PeachDesk shows an estimated cost per minute and latency for the agent as configured, so sourcing decisions are made with numbers in view. Estimates are estimates, not quotes. After real calls, a configurable QA node scores conversations automatically, so quality is measured on production traffic, not guessed from a few tests.
Drafts are free. Publishes are forever.
Edit in draft as long as you like. When you publish, PeachDesk freezes an immutable version of the agent, and every call is pinned to the exact version that served it. Months later, you can open any past run and see precisely what the caller heard, because the definition that produced it never changes underneath you.
Frequently asked questions
Do I need to write code to build a voice agent?
No. Most teams start from one of 26 pre-built templates or describe the agent they want and let the AI Agent Builder generate a starting flow. From there you edit on a visual canvas of typed nodes. A code node and typed Python and TypeScript SDKs exist for teams that want full control, but they are optional.
How do I test a voice agent before customers hear it?
PeachDesk gives every agent four test surfaces: text chat for fast logic checks, AI-simulated caller personas that stress-test the flow, in-browser voice calls that exercise the full speech pipeline, and real phone calls through your connected telephony. Text chat skips the speech stages, so teams confirm voice behaviour in the browser or by phone before publishing.
What happens when I publish a new version of an agent?
Publishing creates an immutable version of the agent. Drafts stay editable, and every call is pinned to the exact published version that served it, so you can always trace what a caller heard. Existing calls are never rewritten, and each version remains inspectable in the run history.
Can an agent use different AI providers for different stages?
Yes. Per agent, you route each pipeline stage, speech-to-text, language model, text-to-speech, or realtime, to PeachDesk-managed providers, your own provider keys, or self-hosted inference, in any combination. The sourcing is set per agent, so one agent can run fully managed while another mixes managed and local stages.