Define once. Run everywhere.
The agent runtime at the heart of Saaya. Wire up identity, knowledge, voice, avatar, tools, and channels in one config, deploy to phone, web, and business messaging from the same definition.
Agents that actually feel like agents.
A Saaya Agent is a single configuration object, prompt, knowledge, tools, voice, avatar, language, and channel, that compiles to a real-time pipeline. The same agent picks up phone calls, joins business messaging threads, and shows up on your website chat bubble. No three-platform Frankenstein.
You version every change. You roll forward and back. You A/B test prompt blocks against each other in production. You watch live transcripts and replays of every session. The boring engineering parts, the orchestration, the model swaps, the retries, the consent flows, Saaya handles. You write the agent.
Under the hood, Saaya is provider-agnostic by design. Leading language models for the brain. Premium neural voices for the voice. Your choice of avatar engine for the face. Switch any layer with a config flag. No rewrites, no lock-in.
Everything you need under the hood.
Agents ships with the boring-but-load-bearing pieces every production deployment needs.
Multimodal in one config
Voice, chat, and video are first-class on every agent. One prompt, three channels, Saaya handles the channel-specific rendering (no bullets read aloud over voice, no awkward silences on chat).
Provider-agnostic stack
Swap LLM, TTS, STT, and avatar providers per agent or per session. No rewrites. Cost-route to a lightweight model for triage, a flagship model for complex calls.
Versioned prompts
Every agent change is a new version. Roll forward, roll back, A/B test individual prompt blocks in production traffic. Audit who changed what, when.
Native tool use
Connect calendars, CRMs, payment systems, internal APIs as tools. Saaya handles function-calling schemas and retry logic out of the box.
English and Hindi
English and Hindi, native-quality. Code-switching supported (Hinglish actually works).
Production-grade safety
Knowledge grounding with a relevance floor you set per knowledge base, and identifier masking on the transcript and call report before they are written. Live audio and the model context are not masked, and there is no consent flow — obtaining consent is yours.
Four steps from blank canvas to production.
A new agent is live in fifteen minutes. Iteration after that is single-file edits.
Define identity
Write the system prompt, pick a voice, pick an avatar (optional), pick a language. XML-tagged template starters for every common use case.
Connect knowledge
Attach one or more Knowledge Bases. The agent searches them mid-conversation and the passages it used come back labelled, with a relevance floor you set so weak matches never reach it.
Wire tools & channels
Add functions for the things the agent should do (book a meeting, refund an order). Pick channels: phone, business messaging, web, video.
Deploy & version
Ship the agent. Watch the session viewer. Iterate the prompt and bump the version. Roll back if a release goes sideways.
A working agent, in fifteen lines.
The same agent definition compiles to phone, web, and business messaging, no per-channel code.
import { createAgent } from '@saaya/sdk'
export const maya = createAgent({
name: 'maya-inbound',
identity: {
role: 'Inbound sales qualifier at Acme.',
objective: 'In under 180s, qualify and book a demo.',
voice: { provider: 'saaya', voiceId: 'default' },
language: 'en-US',
},
knowledge: ['kb_product_docs', 'kb_pricing'],
tools: ['get_calendar_availability', 'book_meeting', 'send_message'],
channels: ['phone', 'messaging', 'web'],
// Per-knowledge-base retrieval tuning. minScore is the relevance floor:
// passages below it are dropped before the model sees them.
retrieval: { minScore: 0.35, topK: 5 },
})On every tier, what you get.
Agents are the core product. Every plan includes the runtime; the difference is volume and provider freedom.
1 production agent · default LLM + voice only · Saaya branding on disclosures.
10 agents · all LLM/TTS providers · Custom branding · Knowledge bases with a tunable relevance floor · Versioning & rollbacks.
DPA · Negotiated agent and knowledge-base limits · Dedicated success manager.
Pair this with the right Solution.
Frequently Asked Questions.
Three things: (1) multimodal in one config, same agent answers phone, chat, and video, no separate builds. (2) Provider-agnostic, swap LLM/TTS/avatar without rewrites. (3) Production discipline baked in, versioning, A/B prompts, strict-citation, audit logs. Most competitors are single-channel and single-provider.
Build your first agent.
Free tier ships a working agent in fifteen minutes. No credit card. Early-bird pricing locked in for life on paid plans.