Skip to main content

​
Browse and deploy voice agent templates

Voice agent templates are pre-built, production-ready voice agents: each ships a tuned system prompt, a Cartesia voice, a curated model, default tool wiring, and a test-call script you can validate in under a minute. Deploying one clones it into your account as a draft agent you configure and promote, just like any agent you author yourself. This is a separate surface from the Agent Marketplace. The marketplace covers text and workflow agents (customer support, sales, scheduling over WhatsApp, email, SMS). The voice template catalog at Agents → Templates (/agents/templates) is voice-specific — every template in it creates a voice-type agent with a real phone voice on its first deploy.

​
1. The catalog

The catalog holds eleven vertical templates as of 2026-08-09, grouped loosely by funnel stage — inbound triage first, then acquisition, operations, collections, feedback, retention, and finally the industry-specific set. Inbound templates (receptionist, retail order status, clinic scheduling, banking card block) are the zero-preparation starters: no contact list to build, just a number to route. Outbound templates target a contact list; the template’s suggested list (for example “abandoned carts, last 72 h”) lands on the deployed agent’s config so the campaign surface can wire a one-click deploy-to-contacts action. If none of the eleven fit, the alternatives are authoring an agent from a prompt — a plain-English brief generates a draft from scratch — or importing a competitor assistant (Vapi, Retell, Bland) you already run.

​
2. Browse and preview

Open Agents → Templates. The catalog renders as a card grid. Each card shows, at a glance:
  • The template’s vertical and category badges (your industry tag plus the funnel stage — sales, support, collections, scheduling, feedback, retention).
  • A one-line summary of the use case.
  • The estimated median call duration, which drives the ROI preview.
  • The number of tools the agent ships wired to.
Click Preview on any card for the full picture before you deploy: the recommended use cases, the complete system prompt, the test-call script (each line is one caller turn you can replay against the deployed draft), and the template’s default model, Cartesia voice + language, and tool list. Nothing in the preview is read-only after deploy — it is all editable on the resulting agent.

​
3. Deploy in one click

Click Deploy on a card (or in the preview dialog). The dashboard calls POST /api/v1/agents/templates/{id}/instantiate with an optional name and description override:
An unknown template id returns 404; a malformed one returns 422 (template ids are lowercase kebab-case). The response lands with 201 and the new agent’s id. The clone stamps everything the template carries into the new agent:
  • System prompt — the full template prompt with placeholders rendered: your organization’s name where the prompt addresses callers on your behalf, and the compliance clauses for your jurisdiction.
  • Model and sampling — the template’s curated model plus its temperature and max-tokens settings.
  • Voice settings — the Cartesia TTS voice id and the voice/STT language.
  • Tools — the template’s default tool wiring.
  • Voice-specific config — the test-call script, the estimated call duration, the suggested outbound contact list, and the resolved jurisdiction’s compliance posture (quiet-hours window, debt-disclosure preroll where required, voicemail redaction policy) as structured configuration the dialer reads directly.
The new agent is always created in draft status — nothing goes live on deploy. The dashboard navigates you to the agent’s studio page to finish configuration; over the API, fetch the agent with the returned id and test via the usual promotion pipeline. Every instantiation is an independent clone. Deploying the same template twice creates two unrelated agents (override name on the second call so the list stays readable), and neither is linked to the other’s edits.

​
4. After you deploy

The clone is production-shaped, not production-finished. In order:
  1. Test it. The draft lands with its test-call script in config — run those turns in the studio, then run the persona simulation and red-team gate before promoting past draft.
  2. Attach a knowledge base. Templates answer from their system prompt alone until you attach knowledge. For anything order- or policy-shaped (retail order status beyond the tracking lookup, receptionist FAQs), build a knowledge base and link it.
  3. Pick a model preset if the default doesn’t fit. The template ships a curated default. Compare the four curated presets on latency, cost per minute, and quality in Pick a voice-agent model preset — the picker applies the bundle to your deployed agent.
  4. Set cost controls. Cap per-run, token, and iteration spending before the draft meets live traffic: Agent cost controls.
  5. Give it a number. An inbound template needs a number routed to the agent — see Set up and run voice queues if callers ever fall through to humans, and Inbound number routing for routing the assigned number itself. An outbound template needs its contact list wired to a campaign before it dials a single contact.

​
5. Template lifecycle

Templates are a curated, versioned catalog — add the ones you want, deploy them, and your copy is yours from then on.
  • Instantiation is a snapshot. The deploy copies the template’s then-current prompt, model, voice, and config into your agent. If the catalog later updates that template — better prompt, different default model — your existing agents do not change. Agents created from a template never auto-update; your draft is fully editable and fully detached, and a template revision can only be picked up by deploying fresh.
  • The catalog is additive. New vertical templates appear as new cards; nothing you have deployed is renamed or retired from under you.
  • The same holds over the API. GET /api/v1/agents/templates lists the catalog (voice templates plus the curated chatbot, router, and workflow starting points the create-agent wizard uses); GET /api/v1/agents/templates/{id} fetches one template’s full payload, including the system prompt and test-call script.
Because a template update never touches deployed agents, treating agent versions and the rollout pipeline as your change management for the deployed copy keeps template upgrades a deliberate, tested decision rather than a background surprise.

​
6. Compliance for outbound voice templates

The outbound templates (cart recovery, appointment reminder, pre-screen, payment collection, lead qualification, survey, win-back) call your contacts; compliance posture on those calls is tenant-owned, the same as any other outbound traffic.
  • Quiet hours. The deployed agent carries a hard prompt rule never to call outside its jurisdiction’s local-time window, plus the matching structured config the dialer enforces. For US destinations this is the 8 AM–9 PM recipient-local federal window described in TCPA federal voice guard; your org-wide posture and per-channel gates are in Quiet hours configuration. Different regions render different windows (EU / UK / MENA / AU / other) from your organization compliance jurisdiction setting — the clone bakes in whichever applies to you at deploy time.
  • Consent and disclosure. Templates that carry a regulator-specific disclosure (the payment-collection template’s debt-collection opening, voicemail handling on the appointment-reminder template) include it in the shipped system prompt for the applicable jurisdiction. You own the consent basis for the contact list you attach — the platform gates dialing windows, it does not supply your consent records.
  • Reachability. The clone inherits nothing about your sending numbers; quiet hours and consent operate on whatever numbers you route to the agent.

​
7. Keep going