Build a voice agent for $0 on Google Cloud free tiers

What you'll need

• Google Cloud free tiers (Cloud Run + Speech-to-Text + Text-to-Speech + Gemini) free tier leanest paid option
  The honest $0 stack: Cloud Run hosts the agent (2M free requests/mo), STT transcribes (60 free min/mo), TTS voices replies (~1M free chars/mo standard voices), Gemini free tier reasons. Billing must be enabled — set a $0.01 budget alert before you deploy.
  Free until: Cloud Run 2M req/mo · STT 60 min/mo · TTS ~1M chars/mo · Gemini free-tier quota. A low-traffic website demo fits inside all four.
  Watch the pricing: Quotas are per-service, not bundled: exceed any one and metered billing starts. A real phone number, PSTN minutes, and SMS are never in the free tier — that's a separate paid channel entirely.
  Pricing checked 2026-10-21 — source.
• Vapi + Twilio (the paid production path)
  When the demo earns a real phone number: Vapi orchestrates the agent, Twilio carries the calls. This is the stack our own receptionist Jordan runs on — per-minute metered, real PSTN.
  Watch the pricing: Vapi ~$0.05/min platform fee plus Twilio telephony (~$1/mo number + per-minute). A busy line is a real monthly bill — which is exactly why the $0 proof-of-concept exists first.
  Pricing checked 2026-10-18 — source.
  Wall check Gates 1/5 ✓ Price ✓ Permission ✓ Gate ! Meter ✓ Exit what we found Voice AI developer API; pay-as-you-go per minute ($0.05/min) with pass-through LLM/telephony costs.

We are not paid by any tool listed here. If a cheaper or better option exists, that's what you'll see.

Pricing above was checked between 2026-10-18 and 2026-10-21. Vendors change their terms without notice, and free tiers do get withdrawn. We take no money from any vendor listed here. If something on this page is out of date, tell us and we will correct it.

How to set it up

1. Open GCP, enable billing, set a $0.01 budget alert first

   Free-tier services still require an active billing account. Before deploying anything, create a budget alert at the lowest threshold — if a quota slips, you hear about it in minutes, not on an invoice. New GCP customers may also qualify for a credit grant; eligibility varies and isn't required for the free tiers.

2. Deploy the agent container to Cloud Run

   The agent is a small websocket/WebRTC service: audio in → STT → LLM → TTS → audio out. Cloud Run's free request allowance covers demo traffic; scale-to-zero means it costs nothing while idle.

3. Wire Speech-to-Text and Text-to-Speech

   STT V1 carries a 60-minute monthly free allowance; standard TTS voices carry a large free character allowance. Stay on standard voices — the premium/neural voices are the paid meters.

4. Point the brain at a free-tier LLM

   Gemini's free tier handles the conversational turn. Keep the system prompt short and the tool list tiny — every extra token per call eats the same free quota faster.

5. Put it on your site as a browser call button — not a phone number

   The $0 version lives behind a 'Talk to the agent' button on your site: visitors speak from their browser, no PSTN involved. The moment you want a real phone number answering, you've graduated to the paid path — and by then the demo has told you whether it's worth it.

6. Watch the three quota meters weekly

   Cloud Run requests, STT minutes, TTS characters — each has its own free ceiling. A quiet demo site stays inside all of them; a promoted launch won't. That's the honest trade: $0 buys you proof, not production.

Common questions

Is $0 really $0, or is there a catch?

The quotas are real and the demo fits inside them — but GCP requires billing enabled, so the catch is discipline, not a hidden fee. Set the budget alert, watch the three meters, and $0 holds. Exceed a quota and metered pricing begins immediately.

Can the $0 agent answer my real business phone number?

No — and that's the line everyone blurs. A real phone number means PSTN, and PSTN is always metered. The $0 build is a browser-side agent on your website. The phone number is the paid upgrade you make after the concept is proven.

What happens when I outgrow the free tier?

Two upgrade paths: pay GCP metered rates as usage grows (cheap at small scale), or move to a managed stack like Vapi + Twilio when you need a phone number, call recording compliance, and uptime guarantees. Both are visible line items — the recipe exists so you can choose with evidence.

Do you run this yourselves?

We publish this recipe because it's the stack we recommend testing first — and we're deploying our own demo build on our Google Cloud credit. Our production receptionist, Jordan, runs on the paid Vapi + Twilio path, which is what a real phone line costs.

Last reviewed 2026-10-21. Tools change — if something here is out of date, tell us and we'll fix it.

Got a different problem?