Translation, documents, projection, and data encoding, with every model running locally. No third-party API calls, no per-token billing, no data leaving the machine you control.
Four applications, one principle. Each one does a job that normally means sending your data to a vendor, and does it on hardware you own instead.
Real-time voice, video, and chat translation where every participant hears and reads only their own language.
Live now and running on the same self-hosted footing as the rest of the suite.
A document editor with word and phrase replacement built into the writing surface, part of a wider productivity set.
Converts text and files into scannable glyphs, then reads them back, including from a photograph of a printed page.
Deploy, share, run. The same three steps for every application in the suite.
A single Linux server. The installer pulls every model it needs once, sets up the service, and takes care of certificates so the browser will trust it.
Nobody installs anything. A link opens the same on a phone and a desktop, which is what makes a call between two countries actually happen.
Audio, text, and files are processed by models on your own server. There is no outbound call to an AI provider to intercept, bill, or log.
Capacity scales with cores, not licences. Roughly two CPU cores carry one live translated call. Adding capacity means a bigger machine, never a higher tier.
Three local models in sequence, each replaceable, none of them a hosted service.
Streaming speech to text that emits sentence boundaries mid-utterance, so the delay does not grow with how long someone talks.
A self-hosted language model held warm in memory. No provider sits in the path, so nothing is rate limited and nothing is metered.
Spoken output in the listener's own language, generated on the same server that did the recognition and the translation.
The first application in the suite to ship. Running now on a single server.
A warehouse manager in Mexico calls a supplier in Seoul. Each one speaks their own native language. Both hear their native language through a real-time translation service.
The browser streams your microphone to the server as you talk.
Recognition, translation, and synthesis run locally, sentence by sentence.
Each participant receives audio and captions in the language they chose.
One address per application. Separate deployments, so a problem in one never reaches another.
| Address | Application | State |
|---|---|---|
| talk.phentropic.ai | Phentropic.Talk, live call translation | Available |
| project.phentropic.ai | Phentropic.Projection | Available |
| draft.phentropic.ai | Phentropic.Draft, document editor | In build |
| glyph.phentropic.ai | Phentropic.Glyph, glyph encoder | In build |
Tell us which application you need and roughly how many people will use it at once. Concurrency determines the hardware, so that one number shapes the whole answer.
Copy this and send it to the address on the right. Nothing has been sent yet.
Deployments are quoted per server, not per seat. If you already have hardware, say what it is and we will tell you how many simultaneous calls it carries.