| Chat, against models running on the machine | in daily use | The everyday case, and the one the rest is built on. |
|---|
| Your documents, read and searched, with citations | in daily use | Retrieval over your own material, on the machine in your building. |
|---|
| Image generation | in daily use | The generator runs on the reference machine and has produced several thousand images; every photograph on this website is its output; the pictures of the machine are rendered from its design file. We do not claim it runs every day. |
|---|
| Video generation | in daily use | Short loops, from a prompt or from a still. The loops on this site are its output. As with images, that is a statement about what it has made, not about how often it runs. |
|---|
| Speech — reading aloud | in daily use | Text to speech, locally. The audio sample on /app/ is the output of this. |
|---|
| Transcription — speech to text | in daily use | Proven by round trip: a sentence spoken by the speech model was transcribed back verbatim in about a second, on this hardware, with nothing leaving it. The model loads on first use rather than at boot, which is why we briefly had this listed as unbuilt — an unloaded model and an absent one look identical from outside. |
|---|
| 3D objects and meshes | in daily use | Twelve distinct meshes ship on this website, twenty-two files once the lighter mobile versions are counted. Eight of them furnish the three rooms you walk through on /journey/; the rest are on the facility pages. Each was generated from a single image and is delivered as GLB. As with images, that is a statement about what it has made, not about how often it runs. |
|---|
| Music | in daily use | Three sub-genres were generated and one seventy-five-second arrangement was chosen for this site. Scored here rather than licensed. Same caveat as the rest: what it has made, not how often it runs. |
|---|
| A coding agent that edits your codebase | in daily use | A workbench on the machine that reads and writes your real files. Every command it runs is checked a part at a time rather than by its first word, anything it cannot parse is refused rather than allowed, and it stops to ask before a write even where its own rules would permit one. What it allowed and what it refused are both recorded, with the rule that decided. It is the newest thing in this table by a wide margin, and two faults were found and fixed in it on the day it shipped. |
|---|
| Standing agents that act to a schedule | in build | The approvals surface is real and in use. Acting without a person in the loop is not. |
|---|
| Garnet Mesh — several machines answering as one | in build | The services are written and installed on the reference machine and have never been started in service. Single-machine Garnet is what works today. Do not buy a second machine for this yet. |
|---|
| Garnet Law | in build | Said on its own page too. We are not selling it. |
|---|