Sunset Mesa Software
Use Ollama to rewrite text in any app
Point Rewire Text at Ollama and every AI transform runs on your own machine: proofread, rewrite, summarize, change tone, translate. Select text in any application, press a hotkey, pick a transform, and the result replaces your selection in place. No API key, no cloud account, and nothing leaves your computer. LM Studio and llama.cpp work the same way.
Why run the AI locally
Three reasons come up again and again. Privacy: the text you transform may be a client email, a performance review, or unreleased code, and with a local model it never crosses the network. Cost: after the one-time model download there is no per-token bill. And availability: a local model works on a plane, on hotel wifi, or behind a firewall that blocks AI endpoints.
The catch with most local setups is plumbing. Ollama gives you a model server and a terminal. Getting from there to "fix the grammar of this paragraph in my email client" usually means writing your own scripts. Rewire Text is that missing piece: a command palette over the selected text in any app, wired to the model server you already run.
Setup with Ollama
- Install Ollama and pull a model. A Gemma-class 12B model is a good default if your GPU or Apple Silicon Mac has the memory for it; 4B-class models fit lighter hardware (see the hardware question below).
- In Rewire Text, open Settings and choose Ollama as the provider. The default server address (
http://localhost:11434) is prefilled, and the app discovers your installed models automatically. - Pick which model handles which tier of work. Rewire Text asks for a capability tier per transform, so you can put a small fast model on grammar passes and a larger one on rewrites.
- Select text in any app, press Cmd+Shift+, or Ctrl+Shift+,, and run a transform. The result streams in from your own machine.
The AI setup guide walks through this in more detail, including hardware expectations.
LM Studio and llama.cpp work too
Prefer LM Studio's model browser and GUI? Choose LM Studio as the provider and Rewire Text
talks to its local server (default http://localhost:1234). Running
llama-server from llama.cpp? Same idea, default
http://localhost:8080. In both cases the app discovers the loaded
models at runtime. And if you use some other runner that speaks the OpenAI API, the custom
endpoint option connects to it with a base URL and a model ID.
What local models handle well
Mechanical language work is where current local models shine, and it happens to be the work you do most often:
Before
The migration have been runing since tuesday and non of the alerts fired, wich suggest the treshold is to high.
After
The migration has been running since Tuesday and none of the alerts fired, which suggests the threshold is too high.
Before
Standup summary: the API rate limiting work is done and deployed, the dashboard redesign is blocked on the new icon set, and QA found two regressions in the export flow that need fixes before Friday.
After
- API rate limiting: done and deployed - Dashboard redesign: blocked on the new icon set - Export flow: two regressions found by QA, fixes needed before Friday
For nuanced tone rewrites, a frontier cloud model still does better. Rewire Text lets you keep both configured and switch providers in Settings, so the local model can be the default and a cloud key the exception.
One more thing that pairs well with local AI: the other half of Rewire Text never needs a model at all. The 43 deterministic transforms (case conversion, encoding, whitespace cleanup, Markdown utilities) are ordinary code that runs instantly on-device, offline, by design.
Transforms for this workflow
- Fix Grammar & Spelling
- Proofread
- Summarize
- Convert to Bullet Points
- Rewrite
- Translate to…
Part of 50+ built-in transforms. See the full list
Questions
Which local model runners does Rewire Text support?
Ollama, LM Studio, and llama.cpp are built in, each with automatic model discovery. Anything else that exposes an OpenAI-compatible server works through the custom endpoint option.
What hardware do I need?
What matters is fast memory for the model: VRAM on a discrete GPU, or unified memory on an Apple Silicon Mac. As a rough guide, 8 GB of either runs a 4B-class model, fine for grammar-level fixes; around 12 GB or more runs a 12B-class model, which handles proofreading, rewrites, and summaries well. Without a suitable GPU, models still run on the CPU, just slowly.
Is a local model as good as a cloud model?
Not at the top end. Frontier cloud models are still better at nuanced rewrites. But mechanical passes (grammar, proofreading, summarizing, bullet points) work well on current local models, and you can mix: local for everyday passes, a cloud key for the occasional heavy rewrite.
Does any of my text reach Sunset Mesa Software servers?
No. There is no account and no telemetry of your text. With a local model, AI transforms never leave your machine at all. With a cloud provider, text goes directly from your machine to that provider.
Does it work offline?
Yes. The 43 deterministic transforms always work offline, and the AI transforms work offline too when they run against a local model.
Try it on your own text
Download Rewire Text and use everything free for 7 days. Buy it once for $29 and it runs on any two of your devices, macOS or Windows.