TensorAgent Desktop

A local AI app for your Mac or Windows PC. Install the app, download a model in Models, and start a conversation. You can chat, work with photos and files, generate images, or use skills to create files and run code. Model capabilities and your computer's memory determine what is available.

Both download buttons open the official release's Assets list. Choose the tensoragent-desktop- file for your platform below. Desktop installers are produced by the updated Release Binaries workflow; older releases may contain only CLI/server archives. If the latest release has no desktop assets, check all releases or use the source build instructions until a desktop release is published.

This guide covers current source. Personal knowledge and response steering were merged after v2026.10.10; a downloaded package needs a newer build containing them. Catalog availability and runtime validation are described below.

1 · Check your computer

The app includes its .NET runtime and native inference libraries; you do not need the .NET SDK, Visual Studio, Xcode or the CUDA Toolkit to install a distribution package. Windows also needs the Microsoft Edge WebView2 Evergreen Runtime; install it if it is absent. The CUDA package carries its user-space CUDA libraries, while the NVIDIA driver must be installed separately.

Plan for at least the 12 GB system-memory tier for the smallest catalog models, and several GB of free disk space for model files in addition to the app. The Models page shows the download size and required memory before you download. Image generation needs the 24 GB tier; larger text and video models need 32–48 GB. These are catalog eligibility checks against system RAM, including on Windows; they do not guarantee that a model fits in GPU VRAM or runs on every backend. Start with Gemma 4 E2B for a smaller first download.

2 · Choose a package

In the release Assets list, V below is the release version without the leading v. Download one package and its matching checksum file.

Checksums are named SHA256SUMS-tensoragent-desktop-V-osx-arm64.txt, SHA256SUMS-tensoragent-desktop-V-win-x64-cpu.txt or SHA256SUMS-tensoragent-desktop-V-win-x64-cuda.txt. To check a download, replace V and the filename in these examples with the version and package you downloaded, then compare the hash with that file's line in the checksum text:

# macOS Terminal, from your Downloads folder
cd ~/Downloads
shasum -a 256 tensoragent-desktop-V-osx-arm64.dmg

# Windows PowerShell, from your Downloads folder
Set-Location "$env:USERPROFILE\Downloads"
Get-FileHash .\tensoragent-desktop-V-win-x64-cpu.msi -Algorithm SHA256

If the hashes differ, download the file again before installing. A checksum verifies the bytes; it does not make an unsigned package signed. Earlier Mac packages were ad hoc signed without Apple notarization. The updated Mac release pipeline requires Developer ID signing and notarization; a corrected download still needs a successful release run with the maintainer's Apple credentials. Existing release assets are unchanged, so check that release's notes and BUILD.txt for its signing status. Windows packages remain unsigned.

3a · Install on macOS

1. Install the app

   DMG: open the disk image, drag TensorAgent.app to Applications, then eject the disk image. PKG: open the package and follow Installer; installation into /Applications may require an administrator password. ZIP: extract it and move TensorAgent.app to /Applications, or to ~/Applications if you want a copy only for your account.

2. Open TensorAgent

   Launch the installed copy from Applications. If an older, unnotarized download shows “Apple could not verify TensorAgent is free of malware”, check the source and checksum. For a copy you trust, try opening it once, then go to System Settings → Privacy & Security, scroll down and choose Open Anyway, then Open in the confirmation. This creates an exception for this app; it does not notarize the existing download. A future signed and notarized release should use the normal downloaded-app confirmation instead. Follow Apple's instructions for opening downloaded apps if the message differs; ask your administrator if the option is unavailable on a managed Mac.

3. Allow the features you use

   Photo, camera, microphone or speech permissions are requested when their features need them. You can review them later in System Settings → Privacy & Security. Ordinary text chat does not require camera or microphone access.

3b · Install on Windows

1. Choose CPU or CUDA

   Choose CPU if you do not have a compatible NVIDIA GPU or want to avoid GPU driver setup. Choose CUDA for an NVIDIA GPU with a compatible driver. Use one installed flavor at a time; the MSI packages belong to the same upgrade family.

2. Install or extract

   MSI: open the package and follow the installer. It installs for the current user under %LOCALAPPDATA%\Programs\TensorAgent and adds a Start menu entry; administrator rights are not required. ZIP: right-click → Extract All into a writable folder, then open TensorAgent.Maui.exe. Keep its DLLs, webui and skills folders together; do not run the executable inside the ZIP.

3. Launch the app

   Open TensorAgent from Start, or launch the extracted executable. Unsigned releases can show a Microsoft Defender SmartScreen prompt. Verify the source and checksum before choosing More info → Run anyway, if your policy allows it. On a managed computer, contact your administrator if that option is unavailable.

If the app reports a missing WebView2 runtime or opens without its chat view, install the Evergreen Runtime and reopen TensorAgent.

4 · Download a model and send your first message

1. Choose language and storage

   Open the app's menu → Settings. Language → App language follows the system by default and offers eight languages. Under Storage → Model download and cache folder, you can set an absolute folder path on a disk with enough room, then choose Save folder. Changing this folder does not move files already downloaded.

2. Open Models

   Use Choose a model on the empty chat, or the menu → Models. Pick Gemma 4 E2B and choose Download. Read the listed model license, download size and memory requirement. A grey Too big entry cannot load on this device.

3. Wait for verification and load

   The progress row shows download and verification. Transfers can resume after interruption. A download selected while you stay on Models loads when ready; otherwise return and choose Use. Wait until the model is in use before sending a message. You do not need to start a separate server or enter an API key.

4. Try a simple chat

   Return to chat and send “Give me three ideas for a weekend project.” Press Enter to send and Shift+Enter for a new line. Use Stop to cancel a reply. The menu lists saved chats; create a new chat when you change topics. Conversations survive app restarts.

Inference runs locally. The first model download needs internet access; after the necessary model files are present, ordinary chat works offline. The Allow network access setting controls model-run code and package installation; it does not prevent the app from downloading the model you requested.

Browse the model catalog

The current catalog has 38 entries. Search by model name, family, quantization or input capability; use All, Compatible and Downloaded filters, expand family groups and keep track of active downloads. Additional entries include larger Gemma/Qwen, GPT OSS, Nemotron, Mistral, Hunyuan, DiffusionGemma, DeepSeek and GLM checkpoints; many remain experimental. Pinned downloads and hash checks establish file identity, while catalog metadata checks do not establish inference on every device. Follow the entry's system-memory tier, inputs and model card; large entries can require far more memory than an ordinary PC.

Steer a reply while it runs

Type and send a new instruction while a reply is generating. The active generation is interrupted, its produced reply stays in the chat, and a new turn follows your instruction after the old generation has drained. With an empty composer, the send button shows Stop; with new content it offers steering. Steering does not undo tool actions already executed.

5 · Photos, files, images and video

• Ask about a photo or file: choose + → Photo or File, attach the item, and describe what you need. The loaded model must support the input. For a vision model without its optional projector, use Models → Add vision, then Enable vision. Settings → Downloads → Include optional files includes the projector and available draft head in future downloads.
• Generate an image: load Qwen-Image 2.1 or Qwen-Image 2.1 Turbo and describe the picture. Turbo uses its fixed 8-step, CFG 1 recipe, with AD-Q4_K at the 24 GB tier and Q8_0 at 32 GB; it shares the base model's encoder/VAE files and refuses step-distillation LoRAs. To edit a photo, attach it and say what to change. Select area beside its thumbnail lets you paint the area to edit. After generation, use Compare original or Edit again. Optional Model → LoRA plug-ins adds separate compatible downloads: fourteen on the base entry, four style/edit adapters on Turbo. See the image editing guide.
• Generate a video: load the appropriate MiniMax-H3 catalog entry, describe a short clip, and use that entry's supported attachments. Video needs a GPU backend; the CPU package does not provide a video fallback. Windows audio/video generation is not covered by the recorded desktop app validation. See video model requirements and limits.
• Keep a result: open a generated file's link and save or copy it to a folder you control. Chats retain their transcript, but generated files and uploads live in the app's cache; keep another copy of anything important before clearing caches or removing the app.

Reuse a library in your chats

Open Menu → Knowledge to create a collection, upload files or a recursive folder, or Add chats with available attachments. Check extraction status, then select Use {collection} in this chat before asking; browsing does not select it. New chats start with no collection selected. Retrieved sources links open cited locations and originals. See the knowledge guide for OCR, supported formats, media-model requirements and limits.

6 · Set up code and browser work

Text chat needs no extra developer tools. Desktop code skills use programs installed on your computer: install Python 3 for Python work and Node.js with npm for JavaScript or the bundled Playwright browser skill. The browser skill requires Node.js 20 or later. Make them available on your PATH, then fully quit and reopen TensorAgent so it discovers them. On Windows use native Windows Node/npm, not WSL.

With a tool-capable chat model, try “Create a CSV of these expenses and check the total with Python.” Attach the input instead of expecting a sandboxed Mac command to read unrelated folders. For browser work, enable the required settings and ask “Use the Playwright skill to visit example.com and summarize the page.” Its first use may download npm packages and a browser. The automated browser has its own session; it does not inherit your everyday browser's sign-in. See browser setup and skill limits.

7 · Update, back up or uninstall

Update: quit TensorAgent, download the new matching platform package from the latest release, verify it, and reinstall. On Mac replace the app in Applications. On Windows run the new MSI; for ZIP installs extract into a fresh folder and use the new copy. Updates are manual. Model files, settings and chats are stored separately from the installed app.

Quit the app before copying these folders. If you selected a custom model folder in Settings, back up or reuse that folder separately. Mac backup tools may omit Caches. Changing the model folder leaves the old files in place; move the model subfolders yourself, with downloads finished and the app closed, if you want to reuse them.

Uninstall: on Mac move TensorAgent.app to Trash; a PKG install has no separate uninstaller. On Windows use Settings → Apps → Installed apps → TensorAgent → Uninstall for MSI, or delete the extracted app folder for ZIP. App removal leaves user data and models available for reinstall. To reclaim individual models, use Models → Delete; Settings → Storage → Delete all chats removes chat history. For complete removal, after backing up and quitting, also delete the data/cache folders above and any custom model folder you no longer need.

The knowledge library lives in knowledge inside the data folder, separately from models/upload caches, and retains imported originals. Quit the app and copy the data folder for backup. Deleting chats or upload caches does not remove imported sources; collection/document removal requires a second inline confirmation.

Troubleshooting

To report a problem, open a GitHub issue with the release version, package flavor, OS, memory/GPU, model and exact error. Check logs under the cache folder's logs directory; review them before sharing because they may contain local paths or prompts.

Validation scope: distribution builds and package checks do not certify every model, GPU, permission dialog or installation path. The recorded Windows desktop app checks cover selected chat, files, code tools, synthetic vision, cancellation and chat recovery; catalog downloads, native pickers, audio/video generation and a long soak were excluded. Mac runs cover specific Apple Silicon models and workflows. See the TensorAgent README for exact device and model coverage; unavailable or skipped scenarios are not passing validation.