Control what it says.
Brands
Your brand voice - without brand risk.
Bonsai gives marketing and CX teams the confidence to deploy customer-facing AI. Structured conversation journeys keep the agent on-brand and on-topic. Built-in guardrails handle sensitive topics, competitor mentions, and edge cases automatically - so your brand reputation is protected by design, not by hope.
-
On-brand by design: Structured flows and approved phrasing keep tone consistent across every channel and locale.
-
Guardrails, not guardropes: Classifiers detect risk in real time. Enforced actions deflect, stop, or escalate - before damage is done.
-
Measurable improvement: Turn real conversations into evaluation scenarios. Ship updates with confidence that quality is trending upward.
Enterprises
Ship AI like software: governed, observable, auditable.
Bonsai treats customer-facing AI operations with the rigor enterprises expect from production software. Role-based access control, environment separation, comprehensive audit trails, and provider-agnostic architecture eliminate the governance gaps that keep compliance teams up at night.
-
Full traceability: Trace every answer end-to-end - from user input through classifiers, guardrails, and prompts to final output. Answer "why did it say that?" with evidence.
-
Controlled rollout: Environment separation (dev / staging / production), approval workflows, and audit logs mean changes ship with accountability.
-
No lock-in: Provider-agnostic architecture supports 13+ LLM, 6+ TTS, and 5+ ASR providers. Swap any time without rebuilding.
Agencies
Ship brand agents faster - without lock-in.
Bonsai gives agencies a repeatable delivery framework for AI agent projects. Reuse proven flow patterns, guardrail configurations, and quality processes across clients - then customize the creative, voice, and integrations for each brand. One framework. Infinite client configurations. No platform dependency.
-
Repeatable delivery: Portable journey and guardrail assets mean every project starts from a proven foundation, not from scratch.
-
Creative + engineering collaboration: Conversation designers shape journeys and tone. Engineers handle tools and integrations. Both work in the same system.
-
Client-by-client flexibility: Optimize provider choices for each client's cost, latency, privacy, and regional requirements. Swap without rewriting.
Developers
The open-source framework for AI agents that are actually ready for production.
Journey-first architecture. Governed by default. Provider-agnostic. Bonsai gives you a clean TypeScript backend, a schema-first API surface, and an opinionated operating model - so you ship production-grade agents instead of perpetual prototypes. Four commands to a running stack. Apache 2.0. Self-host anywhere.
-
Clean DX: TypeScript/Node.js, Express REST API with OpenAPI contracts, WebSocket API with JSON Schema contracts, PostgreSQL with Drizzle ORM, Zod validation, and dependency injection via tsyringe. Modular and event-driven.
-
Provider portability: 13+ LLM, 6+ TTS, 5+ ASR, and 4+ storage providers. Swap per stage or per project - no code rewrite. Adapters for OpenAI, Anthropic, Gemini, Groq, Mistral, DeepSeek, and more.
-
Built-in ops layer: Evaluation tooling, session tracing with full turn pipeline visibility, issue tracking linked to exact events, latency analytics, and immutable audit logs. The hard production stuff is already done.
-
Fast start:
git clone → cd compose → cp env.example .env → docker compose up -d. Console, API, Swagger, and docs - all running locally in minutes.
Every turn stores state, variables, events, artifacts, and timing data - so teams can inspect exactly what happened instead of treating production as a black box.
Personas
Define agent personality, tone of voice, behavioral constraints, and context rules. Every agent starts with a clear identity - not a generic prompt.
Custom Prompts
Build dynamic prompt templates with Handlebars. Inject context, knowledge, and session variables automatically so the agent always has what it needs.
Voice
Configure natural voice output with provider-level controls - voice selection, speaking speed, stability, model choice, and filler responses for perceived-latency management. Support for ElevenLabs, OpenAI, Azure, Deepgram, Cartesia, and Amazon Polly.
LLM 13+
OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, OpenRouter, Together AI, Fireworks AI, Perplexity, Cohere, xAI, and more. Swap per project or per stage.
Text-to-Speech 6+
ElevenLabs, OpenAI, Azure, Deepgram, Cartesia, Amazon Polly. Match the voice to the brand.
Speech-to-Text 5+
Azure, ElevenLabs, Deepgram, AssemblyAI, Speechmatics. Choose for accuracy, latency, or language coverage.
Storage
S3, Azure Blob Storage, Google Cloud Storage, and local filesystem. Keep artifacts where your compliance policies require.
Channels
Web chat, telephony, messenger, kiosks, games, VR/AR, robotics. Headless means any frontend.
Tools & APIs
Sandboxed scripts (isolated-vm), webhooks, MCP support, and LLM-powered tools. Extend capabilities without changing the core.
No rebuild required.
We keep your data safe — learn more
We send updates, not noise.
Built by utter.one, an AI lab at HearMe, with 8+ years engineering conversational AI and a 100% delivery track record.