Coming soon · Windows first

All of your models, one workspace, your machine.

Chat, images, video, music, documents and code, each in its own window, each on a model you choose. Hosted, or running on the graphics card you already own.

Join the list Look inside

No subscription and no credits. Bring your own keys.

Visual Canvas

Shot 01 · three keyframes, one scene

  1. Wide establishing frame: a lone figure on a rocky ridge above a misted valley at golden hour 00:00
  2. The same scene, camera closer: the figure from behind with the valley falling away past their shoulder 00:04
  3. The same scene, camera low: the figure in silhouette against the low sun 00:09
Qwen-Image-2512 Wan 2.2 seed 40218
Code · pi

> npm run build

webpack 5.97 compiled in 8.41s

> git status

3 files changed

Danger guard: rm -rf build/ · allow once?

Council · 3 models 42 tok/s · 18% ctx · cache 94%

Which of these three export paths keeps the timeline frame-accurate?

claude-opus-5 The second one. WebCodecs gives you the presentation timestamp directly, so the cut lands on the frame you saw.
gpt-6-astra Agreed on the timestamp, but the third path survives a dropped frame. Worth the extra pass.
qwen3.5-9b · on your GPU · synthesis
Hold Alt+Shift+Space and talk
Any model, in any window
Hosted or local, they sit in the same list and swap with one click.
Everything lives on your disk
One database file you can copy, back up, or delete.
There is no server to go down
It keeps working whether or not we do.

The desk

Put the work where you want it.

Work runs in parallel: a video rendering in one window while you write in another and a coding agent works in a third. Nothing waits on anything else, and each title bar shows what it is costing you while it runs.

  • Right-click anything and send it somewhere: a render to the canvas, a reply into a document, a file to a coding agent.
  • Zoom out and the whole project sits on one screen; zoom in and you are inside a single conversation.
  • Every action has one name, whether you type it or ask for it.
Council 2 rounds
claude-opus-5 gpt-6-astra qwen3.5-9b · local

Same question to all three. Two rounds of discussion, then synthesise.

Round 2 · lead model writes the synthesis

Models

Point a window at anything.

One live list feeds every picker, and each model carries what it can actually do: vision, tools, reasoning. Your keys are encrypted by the operating system keychain and never reach a window. Claude Code and Codex run on the subscription you already pay for.

  • Anthropic
  • OpenAI
  • Google Gemini
  • xAI
  • Groq
  • Together AI
  • Mistral
  • DeepSeek
  • OpenRouter
  • Ollama
  • llama.cpp
  • LM Studio
  • + any OpenAI-compatible endpoint
Keys
Encrypted by the OS keychain and held in the main process. Nothing on screen can read them. You pay the provider directly; we never see a cent or a token.
Subscriptions
Already paying for Claude Code or Codex? They run on that login and show up in the picker like any other model.
Local engines
Ollama, llama.cpp and LM Studio are servers you run. Umbrelis connects to them and never starts or stops one without being told.
Anything else
Anything that speaks the OpenAI or Anthropic protocol is a form in Preferences: vLLM, SGLang, TGI, or a gateway you host.

Side by side

One question. Every model you have.

Send the same prompt to as many as you like from one composer, then have a model on your own machine grade the answers. Nothing is stopping you comparing a flagship against something free except the tab you would have had to open.

The prompt Our Postgres writes spike to 400ms every Tuesday at 3am. Where would you look first, and what would you rule out?
Model Runs First token Cost Graded Verdict
claude-fable-5-1 hosted 1.9s $0.031 96 Named autovacuum, ruled out locks
qwen3.5-32b FREE your GPU 3.4s $0.000 93 Same answer, asked for the logs first
gpt-6-astra hosted 1.4s $0.024 91 Started outside the database, correctly
deepseek-v4 hosted 2.7s $0.002 88 Lock contention, plausible but second
qwen3.5-9b FREE your GPU 1.1s $0.000 84 Checkpoints. Right family, wrong cause
gemini-3.1-pro hosted 2.2s $0.011 81 Thorough, buried the answer at the end

Illustrative figures, not a benchmark we ran. Your own numbers depend on your card, your keys and your question, which is the point of being able to run it yourself.

Studios

From a sentence to a finished cut.

The Visual Canvas is a board of lanes you wire together: text to image to video to upscale to sound. Each lane runs a local ComfyUI graph or a hosted model, and a prompt writer rewrites your plain sentence into whatever that model wants to hear.

  1. Lane 01

    Prompt writer

    Plain words in, the prompt this engine wants out.

  2. Lane 02

    Image · ComfyUI

    Your own ComfyUI graph, or a hosted model, same list.

  3. Lane 03

    Video

    Turns the frames above into a shot.

  4. Lane 04

    Sound

    Music, effects or narration, generated to length.

  • Studio

    A real timeline. Frame-accurate scrubbing, stacked tracks, an audio mix, and an export that matches the preview.

  • Audio workbench

    Write a song, clone a voice, narrate a script, split stems, clean a take. Every model runs on your machine.

  • Piano roll

    A real piano roll with FL Studio keybinds, quantize, a built-in synth, and your own VST3 and CLAP plugins behind it.

  • Code

    Claude Code, Codex, pi, or Umbrelis's own agent, next to a real terminal. Every command is read before it runs, and the risky ones ask first.

  • Doc

    Opens and saves .docx and PDF, and shows you every AI edit before it lands.

  • Library

    The prompt, the seed and the model stay with the file, so you can make it again.

Consistency

Enough coverage to actually cut with.

Spinning one prompt four hundred times gives you four hundred of the same picture. A shot list gives you something you can edit: several takes of a scene until one holds, then the next scene, with the character, the wardrobe and the light carrying across all of them.

  1. Nine frames from Dawn on the ridge, all the same scene shot different ways

    01 Dawn on the ridge 9 takes

  2. Nine frames from The path, all the same scene shot different ways

    02 The path 9 takes

  3. Nine frames from The shelter, all the same scene shot different ways

    03 The shelter 9 takes

  4. Nine frames from The detail, all the same scene shot different ways

    04 The detail 9 takes

  5. Nine frames from The descent, all the same scene shot different ways

    05 The descent 9 takes

  6. Nine frames from Last light, all the same scene shot different ways

    06 Last light 9 takes

6 scenes
Each one shot until the takes matched each other.
54 takes
Same coat, same light, same weather, across all of them.
$0.00
Every take. The card was already paid for.
$1,340
The same coverage on a metered image API.
Assistant listening

Put the last render on the timeline and score it.

Working in 3 windows
  • Library · picked render_0412.mp4
  • Studio · placed at 00:00:00
  • Canvas · running the score lane

The assistant

One assistant, and it has hands.

There is one assistant and one thread. Reach it from the pill, from Ctrl+K, from your phone, or by holding Alt+Shift+Space and talking. It can see what is open and work in it: run a lane, brief a coding agent, write into another window and wait for the answer.

  • The pill
  • Alt+Shift+Space
  • Ctrl+K
  • Your phone
  • Its own name

First launch

Only what your card can actually run.

It reads your processor, memory, graphics card and free disk, then marks every model as fits, tight or too large for this machine. Install pulls the model, the workflow and its files as one job, and undoes the lot if a step fails.

Store RTX 4080 · 16 GB · 412 GB free
  • Fits Qwen3.5 9B · chat, tools, vision 6.6 GB · Apache-2.0 · runs on 10 GB of card memory Install
  • Fits FLUX.1 schnell · image lane 23.8 GB · Apache-2.0 · ComfyUI graph and its files, as one job Install
  • Tight Wan 2.2 · text to video 28.4 GB · Apache-2.0 · will use most of the card Install
  • Too large Llama 3.3 70B 42 GB · Llama 3.3 · needs 48 GB of card memory Hosted instead
Where the data is
Everything is in one file: umbrelis.db, in your user folder. Conversations, search index, embeddings, window layout, memory and settings. Renders sit beside it and the Library indexes them. Deleting moves to Trash; only you empty it.
Licences
Every model and workflow shows its licence, in the Store and in the Library. Nothing is ever refused on those grounds.

Tell me when it ships

Your graphics card is idle most of the day. Images, video, music, voice and local chat models are what it could be doing instead.

Coming soon · Windows 10 and 11, x64 · No subscription and no credits. Bring your own keys.