Skip to content

Releases: jacobla1/sovatela

Release list

Sovatela 1.2.0

Choose a tag to compare

@jacobla1 jacobla1 released this 10 Aug 21:41
462e649

Sovatela is a desktop chat client for GLM-5.2, Z.ai's open model hosted in France by Scaleway. You bring your own Scaleway key; it is stored in your operating system's credential store and goes only to Scaleway.

Download the installer for your platform below.

Verify your download

SHA256SUMS.txt is attached to this release. Check your file against it before running it:

shasum -a 256 Sovatela_1.2.0_universal.dmg                       # macOS
sha256sum Sovatela_1.2.0_amd64.deb                               # Linux
Get-FileHash .\Sovatela_1.2.0_x64-setup.exe -Algorithm SHA256    # Windows

macOS builds are signed and notarized — verified on this build with spctl, which reports accepted, source=Notarized Developer ID, with the notarization ticket stapled to the app inside the disk image.

Windows and Linux builds are not code-signed — SmartScreen will warn on Windows. Checking the checksum is the verification that matters there.

New in 1.2.0

Terminal access (Settings → Advanced). Use GLM-5.2 from your terminal. It sets up a claude-glm launcher that runs Claude Code against Scaleway through a small local proxy, while your normal claude command keeps using Anthropic models. It reads the Scaleway key you already saved, so there is nothing extra to paste. Claude Code itself is the one prerequisite, and the section checks for it before enabling the button.

It was hidden for the 1.0 and 1.1 releases because running Claude Code against a non-Anthropic model sits outside what Anthropic supports — their gateway documentation says it "doesn't support routing Claude Code to non-Claude models through any gateway". That is unchanged, and it is a support statement rather than a prohibition. What changed is that the caveats are now in the app, next to the button, instead of only in a README nobody reads first.

Two of them matter before you press it. Terminal access is less sovereign than the app: your prompts follow the local proxy to Scaleway, but Claude Code is an agent — it runs commands, installs packages, fetches pages and talks to MCP servers, and those can reach hosts outside Europe. And setup itself reaches outside Europe: it fetches an installer from astral.sh and installs a proxy from PyPI, both US-hosted. Nothing is installed until you press the button, and Settings → Uninstalling & your data removes every piece.

Text size is adjustable (Settings → Appearance), up to 150%. The type scale was written in fixed pixels, so a reader who needed larger text had no way to get it — the accessibility statement called this the most serious gap in the product. The scale is now relative, which makes it responsive, and the app carries its own control, which makes that reachable: a desktop app has no address bar to zoom with, and neither macOS nor Windows passes its text-size setting through to one.

The smallest text in the app was 10px. It is now 11px, and nothing renders below that.

The reduced-motion setting is respected. With it on, animation and transitions stop.

The two status dots that pulse to mean "in progress" are not simply frozen. "Checking" and "offline" are the same colour, and the pulse was the only thing telling them apart — stopping it would have deleted a state for exactly the readers the setting is meant to help. Both dots become unfilled rings instead, so indeterminate still reads as different from settled.

Known limitations

Terminal access on Windows is unverified on hardware. Setting it up has been tested end to end on macOS, and its installer script on Linux. On Windows the app side is confirmed on real hardware — the section appears, Claude Code is detected, and your Scaleway key is read back out of Credential Manager. Pressing Install is the part nobody has run: fetching uv, installing the proxy, editing your user Path, and whether the claude-glm command resolves afterwards. The launcher reads your key by a different route than the app does, so that is unconfirmed too. If it fails, Settings → Uninstalling & your data removes everything it wrote.

Accessibility. Text scales to 150%, not 200%, and spacing does not scale with it, so larger settings tighten the layout rather than growing it — improved rather than resolved. ARIA structure is sparse; there are no keyboard shortcuts beyond Enter and Esc; focus is not explicitly managed when panels open and close; no screen-reader testing has been performed; colour contrast has not been formally measured. Full detail and the remediation plan: accessibility statement.

Not built yet. No conversation export or import (history is plain JSON in a folder you control, so copying it works). No search across chat history. No message editing or regeneration. No automatic updates. No model or provider selection — GLM-5.2 with automatic vision routing.

Model and providers. GLM-5.2's knowledge cutoff is around early-to-mid 2025; turn on web search for current information. It is capable but not frontier for agentic work. Self-hosted SearXNG is fragile under research load. Qwant Staan sign-up is business-only. Web page fetching cannot read JavaScript-rendered pages. Token estimates undercount CJK text, so cost estimates for CJK chats read low.

System requirements

Platform Requirement
macOS 10.15 Catalina or later. Universal build — Apple Silicon and Intel.
Windows Windows 10 version 1803 or later, 64-bit. Edge WebView2 runtime (ships with Windows 10/11).
Linux WebKitGTK 4.1 — Ubuntu 22.04+, Debian 12+, Fedora 36+. A running keyring daemon (GNOME Keyring or KWallet) for key storage.
All An internet connection, and a Scaleway account with an API key. Scaleway bills you directly.

Terminal access additionally needs Claude Code already installed, and on Linux libsecret-tools.

Roughly 100 MB of disk for the application; conversation storage grows with use.

Upgrading

Install over the top. Settings, chat history, memory, projects and stored keys are preserved — there is no migration step and nothing to back up first. On macOS, replace the app in Applications.

Downgrading works the same way: the data format is unchanged across 1.0.0, 1.1.x and 1.2.0. Terminal access is set up outside the app's data folder, so downgrading does not remove it.

Support

Bugs, questions, security reports, privacy and data-deletion requests: info@anaubi.com. Please don't file security issues publicly. Support is best-effort from a one-person project.

Not affiliated with Z.ai, Scaleway, Anthropic, Qwant, Black Forest Labs, Linkup, Mistral or OVHcloud.