Download Catalase
Free for everyone, with no account and no sign-up. Windows and Linux. macOS is not available yet. Setup takes about three minutes, most of which is pulling a model.
Downloads open Monday 16 November 2026. The installer is being prepared. Come back on Monday and the download buttons will be live. If you arrived early, bookmark this page and return then.
Before you download: this is an alpha release. The Windows installer will be code-signed before the production release. See the known limits for a full list of what's still pending.
Windows installer
Windows x64. Updates itself once installed.
Catalase-Setup-0.0.6-alpha.2.exe
Linux AppImage or .deb
AppImage is recommended — it updates itself. A .deb cannot update itself.
Catalase-0.0.6-alpha.2.AppImagecatalase-desktop_0.0.6-alpha.2_amd64.deb
Prefer to build it yourself? The client code is Apache-2.0 and the repository includes Windows, Linux and macOS packaging scripts.
Is there a macOS version?
No. macOS is not available yet. Windows and Linux builds are published. A macOS packaging script exists in the repository, so you can build from source, but no macOS binary is published and the platform has not been tested. If you need macOS today, the comparison page lists which alternatives ship one.
What you need
| Requirement | Detail |
|---|---|
| Operating system | Windows x64, or a Linux x64 distribution that runs AppImages or accepts a .deb |
| Ollama | Needed for local models and inline suggestions. Free and open-source (MIT), available for Windows and Linux |
| Disk space for a model | About 1 GB for the default qwen2.5-coder:1.5b; larger models need more |
| GPU | Optional. The default model runs on a laptop without one — see the latency table below |
| Node.js | Only if you want to use a subscription CLI, which is installed with npm |
| Account | None. There is no login anywhere in the product |
Verify what you downloaded
The release publishes a SHA256SUMS.txt alongside the binaries. Checking the hash is worth the ten seconds, especially for an unsigned installer.
Windows (PowerShell):
(Get-FileHash "Catalase-Setup-0.0.6-alpha.2.exe" -Algorithm SHA256).HashLinux:
sha256sum -c SHA256SUMS.txtCompare the result with the line for your file in SHA256SUMS.txt, published alongside the binaries.
Operating system warnings
The production release will be code-signed for Windows. The current alpha build (0.0.6-alpha.2) is unsigned while the signing certificate is being provisioned.
- Windows (alpha only) — SmartScreen shows "Windows protected your PC". Choose More info, then Run anyway. This prompt will disappear once the production installer carries a valid signature.
- Linux — your distribution may warn about an unverified source. Mark the AppImage executable with
chmod +x.
Always verify the checksum above before running any installer you download from the internet.
Setup, in about three minutes
Install Ollama and pull a model
Get Ollama from
ollama.com/download, then runollama pull qwen2.5-coder:1.5b. That's about 1 GB and runs without a GPU. Roughly two minutes.Install Catalase
Run the Windows installer, or make the AppImage executable with
chmod +x Catalase-*.AppImage, or install the .deb withsudo apt install ./catalase-desktop_*.deb.Open a folder
Choose Open Folder in the sidebar. Catalase indexes the project with Tree-sitter so chat can resolve
file.py:42references to real locations.Start typing
Ghost-text suggestions appear as you type. Tab accepts, Esc dismisses, Alt+] and Alt+[ step through alternatives.
That is the whole setup. Cloud keys and subscriptions are optional and can wait.
What inline suggestions will feel like on your hardware
Catalase tells you which case you are in, in Settings → Editor.
| Hardware | Expect |
|---|---|
| Discrete GPU | 50–200 ms — feels instant |
| Integrated GPU | 300–800 ms — a brief pause |
| CPU only | 500 ms–3 s — behind your typing |
If it feels slow on CPU, pull a smaller model or turn suggestions off in Settings → Editor.
Optional: use a cloud model
Two routes, depending on what you already pay for. Neither is required.
A subscription you already have
Settings → Coding Plans (Subscription). Catalase shells out to the vendor's own CLI, which you sign into yourself — no credential passes through Catalase, and none is stored by it.
| Plan | CLI to install |
|---|---|
| Claude Pro/Max | npm install -g @anthropic-ai/claude-code |
| GitHub Copilot | npm install -g @github/copilot |
| ChatGPT Plus/Pro | npm install -g @openai/codex |
| Google AI Pro/Ultra | npm install -g @google/gemini-cli |
| Alibaba Coding Plan | npm install -g @qwen-code/qwen-code |
| Amazon Q Developer | AWS installer |
| opencode (bring your own plan) | npm install -g opencode-ai |
Sign in once in a terminal, then pick the model from the selector. These show $0.00 (subscription) in the cost meter — not a claim the usage is free, but a statement that billing happens on the vendor's side, so Catalase has no basis to meter it.
Cursor and Windsurf aren't listed, deliberately. Their subscriptions are tied to their own editors, with no CLI a third-party tool can drive. Listing them would promise access that doesn't work.
An API key
Settings → API Keys. Anthropic, OpenAI, Google, Zhipu and OpenRouter. Test verifies the key against the real API before you rely on it. Keys are encrypted at rest with your OS keychain through Electron's safeStorage and never written to a config file.
Coding-plan keys can be served from a different host. Z.ai Coding Plan keys are served from a different host than pay-as-you-go keys. Catalase defaults to the coding-plan host; set ZHIPU_API_BASE if you have a metered key instead. If a key tests as invalid but you're sure it's right, check the base URL first.
Updates
The Windows installer and the Linux AppImage can update themselves. A .deb cannot update itself — reinstall a newer package to upgrade. Upgrading keeps your keys, preferences and conversations.
If something doesn't work
No suggestions. Is Ollama running? curl localhost:11434/api/tags. Is the model pulled? ollama list. Suggestions fail silently by design — they fire on every typing pause, so an error dialog per keystroke would be worse than nothing.
"Model not found." Run ollama pull <name>, or install it from the Marketplace.
A model shows "Not recommended." Your machine's total RAM is below that model's minimum. "Will run, slower" means it fits but memory is tight right now — closing other apps helps.
Chat says the backend is unreachable. The AI backend did not start. Restart Catalase; your files and conversations are kept. If it keeps failing, reinstall from the download page.
More in the user guide and the known limits.