A 27-billion-parameter Jamaican Patois-speaking assistant on CorX Labs' own hardware. It can search the web, run Python, make and edit files, and plan a task on its own, right here in your browser — no toggle to switch on, it just calls what it needs.
CorX3.8 is an open research model and can make mistakes. It searches and runs code on its own when it needs to — every command and search runs for real, watch it in the panel.
Choose a model backend. Bring your own key for a hosted provider, or point at your own CorX3.8 server. Everything is saved in this browser only — each request goes straight from your browser to the provider you pick. New to this? Read the set-up guide.
/v1 is added for you, and it stays saved across a refresh.
Bring a second model into the chat as a colleague. It shares the same
sandbox and the same /work files, has every tool the first one has, and the two of
them hand work over, ask each other questions and review each other in front of you. Useful
when a job really has two halves; overkill for a one-liner.
Paste a GitHub personal access token and the agent can read your repos, create and edit files, commit (push) and open or merge pull requests. The token is stored in this browser only and sent only to GitHub. Make a fine-grained token at github.com/settings/tokens with Contents: read & write (and Administration if it should create repos).
owner/repo or a GitHub URL.
Runs TriStream-SVS on your own GPU. Give it a voice to clone and a performance to sing, and it returns the performance in that voice.
One file. It installs its own dependencies, downloads the TriStream weights and opens a Cloudflare tunnel. You need a machine with an NVIDIA GPU — your own, or a notebook runtime.
Save it as tristream-server.py and run the command below. In a Colab or
molab notebook, paste it into one cell and run that instead — the same file works
either way.
python3 tristream-server.py
First run takes a few minutes — it is downloading PyTorch and the
weights. On Blackwell cards (RTX PRO 6000, RTX 50-series) it installs the CUDA 12.8
build, the first one with sm_120 kernels.
It finishes by printing a https://….trycloudflare.com address.
Put that in the box below and press Connect. Stuck? The
set-up
guide has the same code with the whole thing explained.
A clean clip of the singer you want it to sound like. The speaker encoder takes a voice fingerprint from this and nothing else.
No file chosenThe melody, timing and delivery to reproduce. Only its pitch contour and texture are used — the source stream cannot carry timbre, so this singer's voice does not come across with it.
No file chosenWhat it should sing. The duration and pitch predictors invent the timing and melody from these unless you add a melody guide.
No file chosenAudio in MP3, WAV, FLAC, OGG/Opus or AIFF. M4A and AAC usually will not decode — that is what a phone voice memo records, so export or convert it first. A few seconds of clean singing works better than a whole track; anything over 60 seconds is trimmed.
Saved in this browser only — never sent anywhere. Used to label your messages and, if you turn memory off, kept out of what the agent can recall.
Give it a character and it stops answering like an assistant — it gets opinions, moods, its own way of talking, and can be as short or blunt as it likes. Tools still run for real; only the voice changes.