The Orga platform
Agentsthatsee,hear,andspeak
Real-time multimodal agents that read the camera, understand speech, and reply in a voice that sounds human. Plus a standalone text-to-speech engine in 40+ languages, on one API.
One model. Every dimension of voice.
The same Orga Alondra core powers expression, control, language and sight — so your product never stitches four vendors together.
It laughs, whispers, and pauses to think.
Orga Alondra delivers speech with real intonation, breath and emotion — the uncanny valley, crossed. Listeners can't tell it isn't a person.
Explore the API[warmly] Of course — let me check that for you. <break/> [laughs] found it!
Direct every nuance with a single tag.
Inline control tags steer tone, pacing and emotion on the fly — from calm support to high-energy narration — no re-recording, no fine-tuning.
Explore the APINative in 40+ languages, not translated.
Each language is spoken with local rhythm and accent — so a Spanish line sounds Spanish, not English read aloud with a filter on top.
Explore the APIIt sees what it's talking about.
Feed images and live video alongside audio. Orga Alondra grounds its responses in what it perceives — context-aware voice, end to end.
Explore the APIThe full loop
One agent, every capability
Vision, speech, expression, actions, telephony and analytics on a single API — nothing stitched together, nothing bolted on.
Vision
See
Read the camera, a shared screen, or an image and reason about what is there.
Speech
Hear
Catch tone, hesitation, and interruptions in streaming speech, not just the words.
Expression
Speak
Reply in an expressive voice that laughs, pauses, and shifts tone. First audio in about 95 ms.
Actions
Act
Call your own tools and functions mid-conversation to get real work done.
Telephony
Call
Answer and place phone calls, route them, and deflect, over SIP and the major providers.
Analytics
Measure
Live latency, token, and error analytics for every agent and voice you run.
~95ms
Time to first audio, streamed over WebSocket, before the sentence is even finished.
Powers every kind of agent
Customer support
Resolves queries, manages orders, prepares tickets in real time.
Insurance & banking
Assesses claims, answers questions, automates processing.
E-commerce
Voice shopping assistant that sees the product and responds instantly.
Health & wellness
Proactive support that listens, understands and acts.
What's new
Shipping every week
New releases, model upgrades, and bug fixes — the current state of the platform.
- 2.4.0Latest
Realtime interruption and barge-in
New releaseOct 8, 2026
Voice agents can now be interrupted mid-response, with a tuned end-of-turn detector that cuts the previous reply within 80 ms.
New- Barge-in on the Realtime SDK — agents stop talking the instant a user starts.
- End-of-turn detector exposed as `vad.aggressiveness` so you can tune for support vs. narration.
Improved- First-audio latency on long responses down from 140 ms to 95 ms median.
Fixed- Reconnect loop on flaky WebSocket proxies.
- 2.3.2Sep 20, 2026
Vision grounding in Realtime
Send a video frame alongside audio and the model answers against what it sees — useful for support flows that walk through a UI.
New- `frames` parameter on `/v1/realtime/session` accepts base64 JPEGs or a signed URL.
Improved- Spanish and Portuguese voices get new breath and pause profiles.
Fixed- Audio cuts when the client switches input device mid-session.
- 2.3.0Sep 4, 2026
Orga Alondra 1.5
New model with stronger prosody, lower hallucination on numbers, and a 30% latency improvement on the `/v1/speak` endpoint.
New- Alondra 1.5 is the default for new accounts; opt out per request with `model: alondra-1`.
- Control tag `<laugh/>` now works in Spanish, French, and German.
Improved- Streaming chunk size doubled — fewer network events at the same audio quality.
Fixed- SSML `<break/>` was being dropped on long inputs.
Build with Orga today
Get an API key and send your first request in minutes.