The AI app that learns from your work.
Dark Energy is a Mac app built around an AI model of its own. Chat with it, let its agent search, draft and edit in your folders, find anything in your documents by meaning, and bring in your mail, notes and calendars. Then train its model on your own writing and code. All of it on your Mac.
Everything you do with your work, in one place.
Dark Energy brings chat, an agent, search, classification and your own data sources together around a model that runs on your Mac.
Chat
Ask anything, with your notes and your documents to draw on. Chats are saved, and an answer you rate up teaches the next training cycle.
An agent that does the work
It searches, reads and writes in your work folder, runs commands and checks its own results. Every step is shown, and Undo puts changed files back.
Training
Trains its model on folders of your writing and code, a cycle at a time. Pause whenever you like; it carries on from the same step.
SearchX
Search every document in your folders by meaning and by words, PDFs included, along with the sources they cite.
Classify
See what each document is and how much of it has been trained, and decide what training reads next.
Documents
Browse your documents, filter them down to a regular expression, and get a summary, readability and suggested improvements for each.
Cycles and Introspect
What each training cycle read, how its loss moved and whether it earned its place, and the very row being trained right now.
Data Sources
Bring in Mail, Notes, Messages, Contacts, Calendar, Reminders and Photos if you choose, and have it add the deadlines and reminders you approve.
Menu bar and Shortcuts
Training’s progress in the menu bar, and Shortcuts and Siri actions to start or pause it, check on it, classify a document or search your documents.
It studies your work, and then it does some.
Most assistants know the internet and nothing of yours. Dark Energy’s model starts out capable and general, and keeps learning from the folders you point it at.
Trains on what you choose
Choose folders of your own writing and code. Dark Energy trains on the text, Markdown, LaTeX and source files in them, and puts each finished cycle to work straight away. PDFs are searched, never trained on.
- Pause at any moment; it carries on from the same step
- Pausing frees the memory, so you can chat in the meantime
- Checks each new cycle while the Mac is idle, and stops the moment you’re back
- Measures every cycle on files it never trains on, so you can see what it learned
An agent with real tools
Ask for something and its agent searches your documents, reads and writes in your work folder, runs commands and checks its own work before it answers.
git, Python and Swift type-checking, in your work folder- Drafts documents in plain text, Markdown or LaTeX, and saves them where you can find them
- Shows every step it takes and every file it changed, with an Undo that puts them back
- Keeps notes about you and your work, and uses them next time
Your work never becomes anyone else’s training data.
There is no account and no server of ours. The model, the training and everything it learns live on your Mac.
Runs on this Mac
Chat, the agent and training all run on Apple silicon’s GPU. Nothing you type, and nothing it reads, is sent to a server to be answered.
Only the folders you choose
Dark Energy lives in the App Sandbox. macOS shows it the folders you pick and nothing else, and you can change them at any time in Settings.
Delivered by Apple
The model downloads from Apple as part of the app: about 4 GB to chat, and about 4 GB more the first time you train.
The web, with a switch
Web search and fetching a page are the only things its agent sends out on its own. One switch in Settings turns both off.
Data Sources only on request
Mail, Notes, Messages, contacts, calendars and photos are read only if you import them, and macOS asks you first. Skip them and it never asks.
Bring your Claude Code notes
If you use Claude Code, Dark Energy can bring in the notes it keeps about you, leaving out any that hold passwords or keys.
Four steps, then it gets to work.
The first launch asks for each one. All of them can be changed later in Settings › Folders & Models.
Choose what it learns from
Folders of your own writing and code. You can skip this, and Dark Energy is the model as it comes until you add some.
Choose where it works
The folder its agent reads and writes in, and where it saves what it drafts. It can’t reach anything outside the folders you choose.
Download the model
About 4 GB, from Apple. You can start using the app while it downloads.
Train, and chat
Start a training cycle and keep working. On a Mac with 16 GB of memory, pause training when you want to chat; it picks up where it stopped.
Before you ask
Is Dark Energy an app or a model?
Both. Dark Energy is the Mac app — chat, the agent, SearchX, Classify, Documents, Data Sources and training — and the AI model inside it, which the app runs and trains on your Mac.
Does anything I give it leave my Mac?
No. The model runs on the Mac, and what it learns is stored there. The only
network traffic is the model download from Apple, web searches and pages its
agent opens (which you can switch off), and commands it runs that use the
network themselves, such as git pull in your work folder.
Which Macs does it run on?
Macs with Apple silicon, running macOS 27 or later. 16 GB of memory is enough to train and to chat, one at a time. Chatting while a cycle trains needs 32 GB or more.
How long does a training cycle take?
On a MacBook Air with 16 GB of memory, about 45 minutes for a 400-step cycle. Macs with more GPU cores are faster. A cycle can be paused at any time and carries on from the same step.
What does it learn from?
Text, Markdown, LaTeX and source code in the folders you choose. PDFs are indexed so its agent can search them, but they are never trained on. Build products, dependencies and version-control folders are skipped.
What can its agent run?
Commands in your work folder, in the app’s sandbox:
git, python3, Swift type-checking and the standard
command-line tools, which come with Apple’s Command Line Tools. Mac App
Store apps can’t run programs they have just compiled, so it checks Swift
code with the compiler rather than building and running it.
Do I need Claude Code?
No. If you use it, Dark Energy can bring in the notes Claude Code keeps about you, and learn from your Claude Code sessions how an agent works through a task.
What it needs
- macOS
- macOS 27 or later
- Mac
- Apple silicon (M1 or later)
- Memory
- 16 GB or more; 32 GB to chat while training
- Storage
- About 9 GB for the model and its training copy, downloaded from Apple
- Model
- 8 billion parameters, fine-tuned on your Mac with LoRA
- Account
- None