Genesis Grid
A one-click local AI desktop. Genesis Grid scans your machine, tells you in plain language what it can run, and manages Ollama models for you — no parameter counts or quant jargon unless you go looking for them.
Stack
- Tauri (Rust backend) — hardware detection via
sysinfo, GPU/VRAM probing vianvidia-smiwhere available, and process management for a locally installed Ollama. - React + TypeScript frontend, two themes:
- Light — warm cream, in the spirit of Claude's own interface.
- Dark — "Genesis Grid Labs" neon cyberpunk, pulled straight from
the brand banner (circuit cyan
#2be9ff, neon magenta#ff2ec8, ultraviolet#9b3ffbon void black#05060d).
Features
- Dynamic Model Swarming: Instantly swaps tiny, highly-specialized local models in and out of GPU memory in milliseconds depending on the intent of your query (SQL, Python, Math, Writing).
- Git-Psychology Engine: Reads your past git commits to build a psychological profile of your coding style, perfectly mimicking your architecture and variables so AI code is undetectable.
- Genesis Eye: A native Rust-powered screen capture vision system that gives the AI real-time context of what you are looking at.
- Predictive Shadow Execution: Anticipates your prompts while you are typing and pre-computes responses silently in the background for zero-latency answers.
- Loop Agents: Autonomous AI agents (Architect, Oracle, Cipher) that work together to solve complex tasks.
- Kanban Board: A visual project manager that saves state locally. Loop agents can automatically create and move tasks across the board!
- Notebook Sandbox: A Jupyter-style interface for executing Python/Node.js code locally. The Loop Agents use this sandbox to autonomously write and test their own code.
- Memory Core: A local Markdown-based vault that acts as a second brain. Visualizes your notes as glowing orbs and lines!
Getting set up (development)
npm install
npm run tauri dev
This launches the app with hot reload. The Rust side needs a normal
Tauri toolchain (Rust + platform build tools) — see the
Tauri prerequisites guide
if cargo or a linker is missing.
Getting a one-click .exe (no local Rust/Node needed)
This repo includes .github/workflows/build.yml, which builds real
installers — Windows .exe/.msi, macOS .dmg, Linux .deb/.AppImage
— on GitHub's own servers. You don't need Node or Rust installed on your
own machine for this path; GitHub's runner installs them fresh each time.
- Push this project to a GitHub repo.
- Tag a release:
git tag v0.1.0 && git push --tags - Go to the repo's Actions tab and watch the "Build Genesis Grid installers" workflow run (a few minutes per platform).
- When it finishes, check the Releases page — it publishes a draft
release with the Windows
.exe/.msi, macOS.dmg, and Linux installers attached, ready to hand to your friends.
You can also trigger a build anytime without a new tag, from the Actions tab → "Build Genesis Grid installers" → "Run workflow".
Troubleshooting: Antivirus False Positives (Windows)
Because Genesis Grid is a brand new, indie desktop application, Windows SmartScreen and third-party antivirus software (like Malwarebytes) may flag the .exe as suspicious and try to block the download or installation.
This is a false positive that occurs because the executable is not digitally signed with an expensive Extended Validation (EV) certificate, so Windows does not recognize its "reputation" yet.
To install safely:
- If Windows SmartScreen blocks the app, click "More info" and then "Run anyway".
- If your antivirus blocks it, you may need to temporarily whitelist the
.exefile. All source code is fully transparent in this repository for your peace of mind.
Troubleshooting: Linux (Fedora / RHEL / openSUSE)
If you are running Fedora (39+), installing the .rpm via dnf or KDE Discover may report a missing libwebkit2gtk-4.0 dependency.
Recommended Solution:
Use the .AppImage binary provided in the release instead of the .rpm. It runs self-contained without system package conflicts:
chmod +x genesis-grid*.AppImage
./genesis-grid*.AppImage
(Note: If your system doesn't have FUSE installed for AppImages: sudo dnf install fuse)
Alternative (Installing the .rpm directly on Fedora): Enable the community WebKit 4.0 package on Fedora before installing the RPM:
sudo dnf copr enable hemanth/webkit2gtk-4.0 -y
sudo dnf install webkit2gtk4.0 -y
sudo dnf install ./genesis-grid-*.x86_64.rpm
Troubleshooting: "App is damaged and can't be opened" (macOS)
This is a very common macOS security error caused by Gatekeeper. When a Mac downloads an app from outside the Mac App Store (and it isn't properly signed or notarized by an Apple Developer), macOS flags it with a "quarantine" attribute and will often claim the app is "damaged" to prevent you from opening it.
You can easily fix this by removing that quarantine flag using the Mac's Terminal. Here is how:
Step 1: Move the app to the Applications folder
If you haven't already, drag the "Genesis Grid" app from the disk image (.dmg) or your Downloads folder into your Applications folder. This makes the next step much easier.
Step 2: Run a Terminal command to remove the quarantine flag
- Open the Terminal app on your Mac (you can find this by pressing
Command + Spaceto open Spotlight search, typing "Terminal", and hitting Enter). - Copy and paste the following command into the Terminal:
xattr -cr /Applications/Genesis\ Grid.app
(Note: If the app is named slightly differently, like GenesisGrid.app without a space, adjust the command accordingly, e.g., xattr -cr /Applications/GenesisGrid.app. If you didn't put it in the Applications folder, you can type xattr -cr with a space at the end, and then drag and drop the Genesis Grid app directly into the Terminal window to auto-fill the correct path).
- Press Enter. (It might ask for your Mac login password. When typing the password, no characters will show up on the screen, which is normal. Just type it and press Enter).
Step 3: Open the app Once the command runs successfully (it usually just drops to a new line without any message), you should be able to go to your Applications folder and open "Genesis Grid" normally!
Building locally instead
If you'd rather build on your own machine (both are free):
- Install Node.js (LTS version)
- Install Rust via rustup — on Windows this also prompts you to install the "C++ build tools" component, which you'll need to accept
npm install
npm run tauri build
Produces a native installer per OS in src-tauri/target/release/bundle/
— this is the "one-click install" artifact you'd hand to a friend.
How model recommendation works
scan_system(Rust) reads total RAM, CPU core count, and — best effort — GPU name and VRAM. NVIDIA detection is reliable vianvidia-smi; AMD/Intel/Apple currently fall back to "CPU only" rather than guessing, since a wrong guess is worse than an honest unknown.classify_tiermaps that into one of four tiers: Light, Balanced, Fast, Power User.- The Model Library screen filters the catalog (
model_catalog) into "recommended" vs. "runs, but slower" based on tier fit — never showing raw GB/parameter numbers in the main view. Full specs live behind the "Show advanced details" toggle.
Model catalog (starting set)
Users can add any other Ollama model by name from the "Add your own"
box — it just runs ollama pull <name>.
Branding & ownership
A few things are already wired in so the app is visibly and provably yours:
- App icon —
src-tauri/icons/has a full set (16 up to 1024px, plus.icoand.icns) built from a circuit-node "G" mark in your palette. This is a solid placeholder; for a pixel-perfect set, runnpx tauri icon path/to/your-logo.pngonce you have a final logo — it's the official Tauri generator and handles platform-specific quirks (like macOS's rounded-corner mask) better than a hand-rolled script. - Bundle identifier —
com.genesisgridlabs.appintauri.conf.jsonis baked into every compiled installer; it's what OS installers and package managers use to identify the app as yours. - Copyright metadata — the installer's file properties now carry
© 2026 Genesis Grid Labs™, and the sidebar footer shows it in the running app too. LICENSEfile — an "all rights reserved" proprietary license naming Genesis Grid Labs as the owner, with commit history as the paper trail for authorship.
What actually stops someone from copying it, roughly in order of effort vs. protection:
- A public commit history is your strongest, free proof of authorship. Push early and often — timestamps on a real GitHub repo are hard to fake and easy to point to if authorship is ever disputed.
- The LICENSE + copyright notices make your terms explicit, so reuse without permission is unambiguously a violation rather than a gray area.
- Code signing (a paid certificate, roughly $100–500/year from a CA) makes Windows/macOS show your verified publisher name instead of an "Unknown publisher" warning. This doesn't stop copying, but it does stop someone else's build from impersonating yours convincingly, and it removes the scary security prompt for your real users.
- A registered trademark on "Genesis Grid" (a legal process, not a code change) is the only thing that gives you enforceable rights over the name itself, separate from the code.
None of this makes copying impossible — no license or signature does, since anyone with the installer can decompile it — but it does establish clear, provable ownership and makes casual reuse legally risky, which is what "stolen work" protection realistically looks like for a solo project.
Chat, Projects, and Recent Chats
Chat is a real multi-session system now, not one thread per model:
- "Chat" in the sidebar opens (or starts) the active conversation.
- "Chats & projects" is the browsable history — every past conversation, with a title auto-generated from its first message, which model it used, and when it was last touched. Group chats under a project (a flat name, "All chats" always shows everything), or start a fresh one from there.
- Each chat is its own file on disk under
genesis-grid/chats/<id>.json; projects live ingenesis-grid/projects.json.
Ollama server status is now checked separately from whether it's
installed — ollama_server_running does a real TCP probe to
:11434 every 5 seconds. If Ollama's installed but quit (tray icon
closed, crashed, etc.), the sidebar shows a "Start Ollama" button
instead of falsely claiming everything's fine.
Known rough edges:
goToChat()'s "create a session with my first installed model" logic readsinstalledModelsfrom React state at click time, which can be one render behind a just-completed model install — usually harmless (worst case it picks "freellmapi" once), but worth knowing.- No project renaming/deletion yet, and no way to move an existing chat into a different project after creation — only at creation time via which project tab was active.
FreeLLMAPI key only entered per-screen, no shared store— fixed:load_freellmapi_settings/save_freellmapi_settingsnow persist the unified key and which underlying model to route to (previously hardcoded to"auto", which is also what made it look like "you can't choose the model") togenesis-grid/freellmapi-settings.json, shared by both the Chat screen and "Use online instead."chat_via_freellmapi's error handling used to say only "Unexpected response shape from FreeLLMAPI" with no detail — real testing showed this hid the actual problem. It now surfaces the real HTTP status, the raw response body, and anyerror.messagefield the proxy reports, instead of a generic message.
Still to build
- Download progress streaming for models.
pull_model(the Model Library screen) still firesollama pulland returns immediately — the Ollama and FreeLLMAPI installers now stream real progress (see below), but individual model downloads don't yet. - AMD/Intel GPU detection. Implemented on Windows via DXGI
adapter enumeration (
IDXGIFactory1::EnumAdapters1+GetDesc1().DedicatedVideoMemory) — the same OS-level interface Task Manager and games use. An earlier attempt read a registry field (HardwareInformation.qwMemorySize) that turned out not to be populated by at least one real AMD driver in testing, so that approach was replaced rather than patched further. Still missing on Linux (would needrocm-smifor AMD) and doesn't add anything new on macOS (Apple Silicon is already handled via unified memory). - Streaming chat replies for the FreeLLMAPI test box (currently waits for the full response rather than showing tokens as they arrive).
FreeLLMAPI — now implemented, self-hosted per person
The "Use online instead" screen sets up Ahmed-M-Abdelkader/freellmapi (MIT licensed) as a genuinely working fallback, not just a stub:
freellmapi_statuschecks for Node.js and git, whether the repo's been cloned, and whether the proxy is already listening on:3001.setup_freellmapiclones the repo, runsnpm install, generates a local encryption key, builds it, and starts it in the background — with live progress in the sidebar-style card, same pattern as the Ollama installer.chat_via_freellmapisends a real request through the running proxy using the person's own unifiedfreellmapi-…key (from the proxy's own dashboard, opened viaopen_freellmapi_dashboard).
The honest caveat: this feature needs Node.js 20+ and git installed on the person's machine — Genesis Grid itself doesn't require them, but FreeLLMAPI is a separate Node project, not something that compiles into Genesis Grid's own binary. The UI checks for both and tells the person plainly what's missing rather than failing silently.
It's also self-hosted per person by design, matching FreeLLMAPI's own docs ("run this for yourself, don't share the endpoint") — each friend runs their own local instance with their own free provider keys (Google, Groq, etc.). Nobody's paying for the hosted freellmapi.co tier, but each person does need to add their own keys once, in the proxy's own dashboard.
Ollama auto-install — now implemented, with real caveats
The sidebar's "Install Ollama" button now actually installs it:
- Windows — downloads the official
OllamaSetup.exewith a live progress bar and launches it. You still click through Ollama's own installer window; Genesis Grid can't silently install on your friend's behalf without their consent, and shouldn't. - macOS — downloads
Ollama-darwin.zip, unzips it, and movesOllama.appinto/Applications— no sudo, no terminal. - Linux — runs Ollama's official install script behind
pkexec, which pops a native graphical permission prompt instead of requiring a terminal. This meanspkexec/polkit must be present (it ships by default on most desktop distros — GNOME, KDE, etc. — but minimal or server installs may not have it, in which case this command will fail and the person needs to install Ollama manually).
None of these are silent, admin-bypassing installs — each still asks for whatever permission the OS normally requires. That's intentional: "one click" should mean "no terminal and no separate download page," not "no consent."
Telegram Bot Bridge & Cloud Credentials
Genesis Grid includes a built-in Telegram Bot Bridge that allows you to message your local or cloud AI models directly from your Telegram app.
1. Setup
- Navigate to the Skills & Tools panel in the app sidebar.
- Enter your Telegram Bot Token (obtainable from @BotFather).
- Enter your Allowed Chat ID to restrict bot access exclusively to you (obtainable from @userinfobot).
- Choose your Target Model:
- Genesis Brain: Routes queries to your active local Ollama models.
- OpenRouter / OpenAI / Google Gemini / FreeLLMAPI: Routes queries directly to cloud models using your stored keys.
- Toggle the bridge ON. Detailed execution and polling status logs will print to your console.
2. Usage Notes & Caveats
- First Question Initialization: The very first message you send to the bot might take a few minutes to respond. This occurs because the backend is initializing connection pools, checking local models, or warming up API endpoints.
- Typing Status Bubble: While the bot is processing or generating a reply, it will display a live "typing..." bubble inside Telegram so you know it is actively working on your answer.
- Simultaneous Cloud Keys: You can configure and save API credentials for OpenRouter, OpenAI, Google Gemini, and FreeLLMAPI concurrently in settings. Dedicated "Save" buttons next to each credential input field persist changes instantly with a green confirmation checkmark.
- OpenRouter Free Models: When using OpenRouter, you can select from a dynamic list of currently available free models directly inside the chat interface dropdown, which fetches model metadata in real-time.