Everything you need to run Refinix.

What you need

A computer you control

Windows, macOS and Linux are all targets. macOS is the strongest desktop build today; Windows and Linux desktop builds are still being qualified on specific hardware. Which models fit depends on memory.

A local model engine

The current build talks to Ollama on this computer over 127.0.0.1, with models pulled separately. The Beta is meant to manage its engine and models for you, without a terminal.

At least one open-weight model

Refinix runs a model only for workflows it is qualified for (tested end to end on named hardware), so installed does not mean supported. The model button marks one that is not available.

No internet connection

Not needed for daily work. Setup uses an explicit download or, as planned, a verified offline bundle from removable media; after that, Refinix is designed to run with the network unplugged.

First run

1. Get the build Download the package for your computer from the downloads, and check it against the SHA-256 published beside it. macOS: open the DMG and drag Refinix to Applications; it is not signed by Apple, so the first open is blocked — choose Open Anyway in System Settings → Privacy & Security. Windows 11: run the setup (it installs for your account); it is not code-signed, so SmartScreen may warn — choose More info → Run anyway. With Smart App Control on, Windows blocks it; we do not ask you to turn that off. Ubuntu 24.04: open the .deb with App Center.
2. Open Refinix and let it look around On first launch it checks this computer for a model engine and reports what it found. The line at the bottom-left of the window is the short version: Ready on this computer means the engine answered and the selected model is eligible for local Chat.
3. Check the engine in Settings Open Settings from the bottom of the sidebar. AI engine and model names the engine it reached, the model in use and any others available. If it says the engine did not answer, start the engine and press Try again.
4. Ask your first question Go back to Chat, type into the box at the bottom and press Enter. The first answer from a cold model usually takes longer. It has to be loaded into memory before it can reply.

The window

Refinix is one window in three columns. Both side columns fold away with the two buttons at either end of the header, so the conversation can have the whole width when you want it to.

The round button in the header switches between light and dark. The icon shows the theme that is on: a moon and star while dark, a sun and cloud while light. The choice is remembered.

Chatting

Type into the box at the bottom and press Enter to send. Shift + Enter starts a new line instead, so you can write a paragraph without sending it half-finished.

• While it is working the mark beside the box turns and the header says the request is running. A Stop button appears next to Send. Pressing it ends that reply and keeps everything already written.
• The model button at the bottom-right of the box names the model answering. Click it to switch to another installed model; one not available for this workflow says so.
• Scrolled up mid-answer? A small arrow appears over the conversation; click it to jump back to the newest text. Refinix will not yank the view down while you are reading.
• Your draft is kept. If you type something and switch conversations, the half-written message is still there when you come back.

Managing conversations

The sidebar lists conversations on this computer: pinned first, then most recently updated.

• + New conversation sits at the top of the list, above your history, so starting fresh never means scrolling.
• The magnifier beside the word Conversations opens a search box. It searches titles and message text; closing it puts the full list back.
• The … on a row opens the actions for that conversation, renaming and pinning among them.

Files and documents

The + button at the bottom-left of the message box attaches files to a request, or picks a document skill. You can also drag files onto the message box.

Reads today

txt, md, csv, json as plain text, docx, xlsx, and pdf with a text layer (selectable text). In Chat, one png/jpg as the only attachment can be described by a vision-capable model, not a qualified document-OCR result.

Writes today

docx and PDF.

Per request

Up to 6 files read, 20 MB each.

Searching documents

Keyword search (SQLite FTS5): it matches the words you type, not their meaning. Semantic search with a local embedding model is planned; the product labels this rather than implying more.

Changing code

The Code surface edits files in a project folder you choose. The order is always the same: pick the folder, pick the files that may change, describe the change, then read the diff.

1. Choose a project Choose project above the message box picks the folder. Paths outside it, including symlink escapes, are refused.
2. Select the files Use + to pick the files that are allowed to change. Refinix sends the minimum set, not the repository.
3. Describe the change and press Propose You are asking for a proposal, not an edit.
4. Read the complete diff, then accept Every line that would change is shown before anything is written to disk. Local changes are marked not sandbox tested and can be reverted with Undo this change.

How much it remembers

A model can only consider so much text at once. That budget is the context window, and Refinix divides it between the conversation it sends and the room it leaves for the reply. You can see the exact split in Settings → Advanced → Conversation limits, drawn as a single bar.

• When a conversation gets long, older complete exchanges are left out of the next request so it still fits.
• Your saved history is never changed. Leaving an exchange out of one request does not delete it. Scroll up and it is still there.
• The meter beside + shows how full the current request is, so a long conversation is visible before it becomes a surprise.

Settings and Advanced

Settings answers one question per card: is this computer ready, is the engine ready, is anything else connected, and is anything running right now.

Advanced is collapsed by default and nothing in it is needed for ordinary use. It opens with three charts of what this computer has actually been doing: how long replies take, which hours it was busy, and how requests ended, followed by the engine address, model settings, the token window, addresses and the recent job record.

Not in this build

Scanned PDFs

PDF text layers are read; scanned pages remain unread without supported OCR.

Scans and photographs of text

Not read as document text: OCR is not yet qualified, and Refinix will not guess at words in a picture. Chat's single image description is not a qualified document-OCR result.

Meaning-based document search

Not available: search matches words, not meaning.

Creating, deleting or renaming files

Not available on the Code surface in any access mode.

Running commands, Git, installing packages

Not available, on this computer or through a paired worker. Separately, a paired worker can run a change's tests in a network-off sandbox. Experimental.

Sharing work with another computer

Pairing and remote jobs exist in early form and are experimental; they are not yet qualified for general use.

Common questions

Our team and contributions

• Aditya: Team lead, backend engineering and deployment. Leads the team and works on the backend and deployment.
• Vedant: UI/UX, frontend engineering and planning. Plans the interface and builds the frontend experience.
• Yug: Model research and routing. Researches local models and works on model-routing decisions.
• Sahil: Research and debugging. Contributes research across the project and helps investigate and fix issues.
• Tanvi: Kubernetes and distributed architecture planning. Plans the Kubernetes and distributed architecture work.
• Prachi: Kubernetes and distributed architecture implementation. Implements the Kubernetes and distributed architecture work.