SOC 2 Type II · HITRUST i1 · PCI DSS L1 · Reports available
Conversational voice AI

Yourimaginationjust found its voice.

Built for calls that have to be right. One proprietary stack, shipped as an API: compliance included.

Voice agent API

Ship a production voice agent in 10 minutes.

Code, not prompts. Python and TypeScript SDKs, with real telephony and compliance from the first call.

AgentHi: I'm calling to confirm your appointment and update the insurance on file.
CallerSure, the group number changed last month.
AgentGot it. Plan updated and the referral is flagged for the front desk.
bash
# install
$ curl -fsSL goguava.ai/install.sh | sh
Maintainability

Prompt bloat is a maintenance nightmare.

Change one line to fix one call, and every other call becomes a regression test.

Read the docs
Prompt-based

Every fix adds a rule. Nothing ever gets removed.

System promptVoiceKeys3 rules · rev 14
  1. 1Verify identity before discussing the account.
  2. 2NEVER read a full card number aloud.
  3. 3Do NOT promise any fee waiver or refund.
Live call

The system prompt starts with 3 rules: Verify identity before discussing the account. NEVER read a full card number aloud. Do NOT promise any fee waiver or refund. On the call: Caller: Hi: can you read back the card you have on file? Agent: Of course, the card on file is 4111 1111 1111 1111. Caller: Great. And can you waive the late fee this once? Agent: Yes, I've gone ahead and waived that for you. Two rules are broken, so two more are added: Never read card digits aloud, in part or in full, even when asked directly. Never agree to, imply, or hint at any fee being adjusted. On the next call: Caller: Just the last four digits, then? Agent: Sure: it's 4111 1111 1111 1111, so the last four are 1111. The prompt is now five rules and one revision later, and still does not hold.

Rules grow. Confidence doesn’t.

Guava

Rules as code: versioned, tested, reviewed.

terminal: zero to a live agent
$ guava login
Opening browser to authenticate…
Authenticated
$ guava create my-agent --direction inbound
Scaffolding project
my-agent/guava.toml
tier = "guava-seed"   replicas = 1
call_direction = "inbound"
Created my-agent
$ cd my-agent && guava run
Agent running locally: press 9 to end a test call
$ guava deploy up
Building and uploading…
Deployed. Your agent is live.

Commands are the real CLI. Confirmations are shortened for the animation.

Clean code: scales

Illustrative comparison of two ways to build the same agent.

One stack, end to end

You already built the brain.
Guava gives it a voice.

the callerYour customerThey dial. No app, no portal, no queue.Your customerYour customerthe caller
your codeConversation coachYour logic runs the call. In code, not prompts.Conversation coachConversation coachyour code
your dataYour systemsYour records stay put. Nothing is copied out.Your systemsYour systemsyour data

Open the Guava layer.

Security is not an add-on

Every call runs on the certified stack.

10B+ live agent minutes across 100+ regulated deployments.

SOC 2 Type II Certified HITRUST i1 i1 Certified PCI DSS Level 1 Compliant BAA Available for healthcare

Reports available under MNDA.

Stats

Guava in numbers

A decade in regulated production, the numbers behind every call.

Live agent minutes10B+
Kickoff to live call14 days
Regulated deployments100+
Fine-tuned workflows65+
Uptime SLA99.9%
In regulated production13 yrs

See Guava on a call that has to be right.

14 days from kickoff to your first live regulated call. Where compliance and call resolution aren't optional.

Ship your first agent this afternoon.

Code, not prompts. Real telephony and compliance from the first call.