Modex Desktop 0.1.1 — public preview

The second public preview. Windows (x64) only.

This build is unsigned, so Windows will show a SmartScreen "unknown publisher" prompt. Verify the installer instead of trusting the prompt:

Get-FileHash Modex-Desktop-0.1.1-x64.exe -Algorithm SHA256

SHA-256	F344C4DFB06318CB39F850671EC89349969235F565D7163E5FD2F20D6E48313A
Size	366,345,936 bytes (349 MB)
Installed footprint	about 1.3 GB

If your antivirus quarantines the download, restore it only after the hash matches. We will never ask you to disable Windows Defender.

Download and full guidance: https://modexai.io/downloads.html

Before you upgrade

• Modex can now install two optional toolchains — a TeX distribution for LaTeX, and a Python environment for analysis. Neither is downloaded unless you ask for it, and each states its size first.
• Load / Unload in Settings → Libraries is now Show / Hide. A library you had loaded is simply shown; nothing is uninstalled.

New

• LaTeX, with the PDF beside the source. A .tex file opens with highlighting and its own Compile button, and compiles on save. Modex can install a pinned TinyTeX (~760 MB) for you, or use a TeX installation you already have.
• A real PDF reader — outline, zoom to the pointer, grab-to-pan, find.
• The agent runs Python. It can run a script against your simulation results and place a figure it computed into a Modex document. Modex can provision a managed environment (~400 MB) after an approval screen, or use an interpreter it finds on your machine.
• Attach an image or a PDF to the agent — paperclip, paste, or drag from the workspace.
• A chat transcript that reads like one. Text, tool calls and figures in the order the model produced them, rendered markdown, clickable file mentions, and a docked session column instead of an overlay.
• Workspace files open as what they are — images, media, markdown, CSV — and the explorer can create any file type.
• New Settings pages: Simulation, OMC, Python and LaTeX.

Improved

• Settings → Libraries is one list of every library Modex can see, each with one row and one eye. Hiding sticks across a reopen.
• Simulation runs no longer accumulate without asking — results have a retention setting, and a run a figure depends on stays protected.
• Every field in the run dialog now reaches the solver, or says it cannot.
• GPT 5.4 and GPT 5.4 Mini are offered again. They are never chosen automatically, so the default is unchanged.
• The update notice can be dismissed.

Fixed

• The assistant can read the files in your workspace again on Windows. In 0.1.0 it refused ordinary files sitting in your workspace root. Nothing was ever at risk — the check was refusing too much, not too little.
• The first simulation of a session no longer fails after 15–20 seconds. A cold compiler is now given its full start-up budget, and is warmed at launch rather than on your first Run.
• Quitting while Modex is still starting no longer leaves a process behind, which could stop the next launch.
• Upgrading no longer removes Modex's registrations from your other AI tools.
• A failed agent run tells you what to do — the action to take comes first, the provider's own output follows.
• The transcript no longer goes blank while the agent is working. A Thinking cue now covers the stretches where the model works before it writes.
• A turn's copy button appears when the turn is over, not while it is still running.
• Your choice about anonymous usage data is now honoured by every part of the app. The switch is in Help → About Modex, alongside the update preference.

Known issues

• Upgrading from 0.1.0 clears the registrations in your other AI tools once. An upgrade runs the installed build's uninstaller, so no new installer can change 0.1.0's behaviour. Modex restores the entries the next time it starts, and upgrades from this release onward keep them.
• The two Install buttons do not go through an HTTP proxy. On a machine that reaches the internet only through a corporate proxy, installing TeX or Python from Settings will fail. Point Modex at an interpreter or a TeX installation you already have (Settings → Python / LaTeX). The OpenModelica install is unaffected.

Modex Desktop 0.1.0 — public preview

First public preview. Windows (x64) only.

Before you install

This build is unsigned, so Windows will show a SmartScreen "unknown publisher" prompt. That is expected for a preview — verify the hash instead of trusting the prompt:

Get-FileHash Modex-Desktop-0.1.0-x64.exe -Algorithm SHA256

SHA-256	EFF167F2C2DC702CC936335AC3323066948656134E35B186121B74CF88C29D4A
Size	350,359,640 bytes (334 MB)
Installed footprint	about 1.3 GB

The download is large and the install is larger — the build carries the supported agent runtimes and a self-contained Modelica server, so there is no separate Node, npm, or server install to do.

If Windows Defender or another AV quarantines or deletes the download, restore it only after the hash matches the value above. We will never ask you to disable Windows Defender.

Full guidance: https://modexai.io/downloads.html

What this is

An Electron desktop host that runs the Modex web app over a self-contained Modelica runtime. The API server is spawned from a bundled Node binary with a pruned production dependency tree.

• Per-user NSIS installer with a chooseable install directory.
• A generated THIRD_PARTY_NOTICES.txt covering every bundled dependency ships inside the install.
• Agent models. Two providers, each driven by its own bundled CLI: Anthropic (Claude Code) and OpenAI (Codex). The OpenAI picker offers what the bundled Codex CLI 0.146.0 advertises — for this build gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5 and gpt-5.2. GPT-5.4 and GPT-5.4 Mini are not among them: OpenAI steers users to GPT-5.6 and hides both from the CLI's catalog, and this build offers only what that catalog lists. The service itself still runs them.
• Update notification — notify only. Modex Desktop checks modexai.io/latest.json at most once a day and tells you when a newer version exists, with Download and What's new. It never downloads or installs an update by itself: the build is unsigned, so nothing would vouch for what it fetched. About → Check for updates automatically turns it off; the manual Check for Updates… keeps working either way. The check sends no usage data, account, or install ID, and carries no version.
• Anonymous usage statistics, on by default. A fixed set of events: that the app launched; the outcome of a startup, runtime, or compiler install attempt; the outcome of a simulation run and of an AI request (the family of provider and model, never the model name you typed); and the name of a public Modelica library a simulated model declares, with anything from your own workspace reported as the single word custom_or_unknown. Never file names, model content, prompts, or error text. Help → Privacy & Telemetry switches it off at any time, and https://modexai.io/privacy.html lists every event, who processes it, and how long it is kept.

Built on Electron 41 (Chromium 146, Node 24).

Known issues

• If you also have a standalone Codex CLI newer than the bundled 0.146.0, the two versions share %USERPROFILE%\.codex and take turns rewriting models_cache.json in shapes the other cannot read. You may see failed to load models cache: missing field supports_reasoning_summaries. It is harmless — the CLI refetches the list — and it is not the cause of a failed run.
• A failed agent run can show raw provider output (for example a stack of 401 Unauthorized lines from wss://api.openai.com) instead of a "sign in" prompt. If a provider's models are listed but every run fails at the vendor, check that you are signed in to that provider's CLI.

Reporting problems

Please include your Modex Desktop version and what you were doing. The in-app support bundle carries the runtime version and capabilities and deliberately contains no workspace path, port, or token.