A desktop assistant that reads your files, looks at your screen when you ask, and writes real documents — running entirely on the computer already on your desk.
Every mainstream AI product is a seat in someone else's datacenter. Your files go up, an answer comes back, and a subscription renews every month. Lemonade runs the other way round.
| Cloud AI | Lemonade | |
|---|---|---|
| Where it runs | Someone else's servers | Your computer |
| Your files | Uploaded | Never leave the machine |
| Cost | Monthly, forever | $49, once |
| Works offline | No | Yes |
| Rate limits | Yes | Your GPU is the limit |
Eleven things it actually does. No model API key or cloud inference upload.
Point it at a folder and it builds a private index. Ask "what changed in the contract since Friday" and it reads your actual documents. Indexing is incremental — a file is only re-read when it changes.
Drag a region, capture a window, or paste an image, and ask about it. The capture stays on the machine like everything else.
Ask for a PDF, Word document, PowerPoint deck or Excel workbook and it produces a genuine one — real formulas in spreadsheets, real slides in decks, openable in Office.
Before writing a file, it shows you the finished document, a line-by-line diff, and an editor. Approve it, or fix its version and save yours instead.
Documents open in a real editor: a spreadsheet grid with live formulas, a slide editor with a filmstrip and speaker notes, or a continuous rich-text page. Saving writes the real file back.
Local models through Ollama, which Lemonade installs and manages itself. No model API key or Lemonade account. Optional mailbox connections sign in directly with Google or Microsoft.
It checks your GPU, VRAM and RAM, then downloads models sized to what you actually have. No terminal, no config files.
Ask a coding question and it routes to a code model; drop in a screenshot and it switches to a vision model. It manages memory itself, unloading one model to make room for another.
Search your whole home folder by filename or by what is inside the files, without indexing anything first.
Reading files, running commands, capturing the screen and searching the web are each Allow / Ask / Never. Nothing happens silently.
Web research and mailbox access are permission-gated. They send the approved query or mail request to the relevant service; model inference and the rest of the conversation stay local.
A graphics card is not required. More VRAM buys bigger, smarter models — it is the upgrade path, not the entry fee.
With 32 GB of system RAM, the optional tiered runtime can stretch a 10 GB GPU to capable sparse models that would not fit in VRAM alone. It uses full-model, quantized weights and does not promise small-model speed.
| OS | Windows 11 or Linux (64-bit) |
|---|---|
| RAM | 8 GB |
| GPU | Not required — runs on CPU |
| Disk | 2.6 GB total |
| OS | Windows 11 or Linux (64-bit) |
|---|---|
| RAM | 16 GB |
| GPU | 8 GB VRAM or more (NVIDIA or AMD) |
| Disk | 5 GB total, more for extra models |
| Lemonade itself | 122 MB |
|---|---|
| Model runtime (once) | 1.4 GB |
| Smallest usable model set | 1.1 GB |
| A good general set | 3.6 GB |
| Your GPU | What you can run | Total disk |
|---|---|---|
| None (CPU) | Small chat model + file search | 2.6 GB |
| 8 GB | Chat, images and tools in one model | 5.1 GB |
| 10–12 GB | Add a dedicated vision model | 11.1 GB |
| 10 GB + 32 GB RAM | Tiered 35B sparse model | about 27 GB |
| 24 GB | Large model, vision and code together | 20.6 GB |
This is the actual reason to buy it, so here it is plainly.
Runtime and model downloads, update checks, licence activation and occasional rechecks, connected mail, approved web research, and any network access performed by a configured MCP server. Lemonade does not upload your documents or conversations to a cloud inference service; an external tool receives only the approved request and data needed to perform that operation.
No. It runs on CPU alone. It is much faster with one, and a bigger card lets you run bigger models — see the table above.
Yes, once your models are downloaded. The initial setup needs a connection to fetch the runtime and the models themselves.
No. $49 once, for a perpetual licence. There is a 14-day free trial first — no card required.
A local model on consumer hardware is not as strong as the largest cloud models, and we are not going to pretend otherwise. What it is very good at is working with your files. It is private, it works offline, and it does not charge you monthly.
Yes. It writes real .docx, .xlsx, .pptx and .pdf files — with working formulas and real slides — and they open in Office.
Windows 11 (64-bit) and Linux x86-64, the latter as an AppImage. A macOS build is not published yet.
There is no account to sign into, so we send the key again to the address you bought with.
No datacenter. No subscription. No upload. Try it free for 14 days — no card — then buy a licence when you are sure.
