Grok Build is open source (Apache-2.0), and Swerve Build is the calm, auditable way to run it on Windows. Scope work to a project, watch every tool call stream through an approval gate, let agents run themselves on a schedule, and record your screen with desktop audio, all in one native app with no browser tab.
Prefer to keep your code on your machine? Point Grok at your own local or self-hosted endpoint, or run local GGUF models. Neither needs an xAI sign-in.
- Project chats. Folder-scoped conversations with ACP streaming, a tool-approval UI, inline images and video, and an honest context bar (real token counts from the agent, or a dash when the agent reports nothing, never invented). Up to 3 run at once.
- Any agent, managed in-app. Grok works out of the box. Claude Code and Gemini install, sign in, and uninstall from Settings.
- Automations. Headless Grok runs on a schedule, a git commit, a file change, or on demand. Read-only by default.
- Workflows. A visual node-graph engine: triggers, requests, transforms, sandboxed code, and agent turns.
- Screen recorder. Real display capture to H.264 MP4 (NVENC, software fallback) with desktop audio.
- Local models. An app-managed
llama-server(pinned and checksummed) with a live VRAM meter. - Opt-in agent surfaces. Let an agent run terminal commands, drive a debug browser, or drive Swerve Build's own UI. Each sits behind its own explicit, off-by-default grant.
A trigger (a schedule, a new commit on a branch, a changed file, or a manual Run now) fires a headless Grok run against a project.
- Safe by default. Every automation starts in shadow mode: a read-only tool set enforced in the app, so a background agent can read and reason but never touch your files. Network tools are off by default too. (Write mode is gated behind a later build.)
- DST-safe schedules, auditable run transcripts under
~/.swervebuild/runs/, and composable runs that chain one rule's output into another's prompt. - Start from a recipe (Doc-drift check, Project summary, Dependency digest) and tweak from there. Automations run on Grok, and the app says so when another provider is active.
Open Settings → Providers. Claude Code and Gemini are installed by the app at pinned versions, signed in via each agent's own flow (a button opens it; Swerve Build never sees your credentials), and removed with a confirmation that names the exact command.
| Provider | How to enable |
|---|---|
| Grok (default) | Install and sign in on the home screen |
| Claude Code | Settings → Providers → Install, then Sign in (claude /login opens in a terminal) |
| Gemini | Settings → Providers → Install, then Sign in (Google OAuth in your browser) |
| Ollama | Already covered. Use Local models (managed llama-server) |
| OpenAI-compatible | Already covered. Use the Grok custom endpoint below |
| Anthropic API (direct) | Not built yet |
No Node.js or Rust required for end users.
- Download the latest Windows installer from Releases (
Swerve Build_<version>_x64-setup.exe). - Run it and open Swerve Build.
- On the home screen, click Install Grok Build (installs the official Grok CLI, pinned and checksum-verified, to
~/.grok/). - Click Sign In and finish OAuth in your browser.
- Open Projects, add a folder, and start a chat, or head to Automations and start from a recipe.
Swerve Build ships unsigned. There is no code-signing certificate, so the first run of a new release shows "Windows protected your PC". Click More info → Run anyway. Your browser may likewise flag the download as "not commonly downloaded". This is expected for every release, because unsigned reputation attaches to each file's hash.
Verify what you downloaded instead of trusting the warning away. Every release publishes a SHA-256 checksum next to the installer:
Get-FileHash '.\Swerve Build_<version>_x64-setup.exe' -Algorithm SHA256Compare against the value on the release page. Only download from GitHub Releases. There is no other distribution channel.
In Settings → Custom endpoint (advanced), point Grok Build at any OpenAI-compatible inference: local (Ollama, llama.cpp), self-hosted, or a gateway. Swerve writes a managed [model.swerve-endpoint] block to ~/.grok/config.toml (backing the file up first) and routes Grok's default model to it while enabled. The endpoint's API key is stored in the Windows Credential Manager, never in a plaintext file and never written into Grok's config.toml, then injected into Grok's environment at launch. Because it satisfies Grok's own auth, no xAI sign-in is required, so your code can stay entirely on your machine. Both chats and Automations follow the endpoint.
Settings → Local models manages a pinned, checksum-verified llama-server: register GGUF files and they appear in every model picker (chats and automations) with a live VRAM meter while loaded. Chats and background automations lease the server cooperatively, so one can't yank a model out from under the other.
Swerve Build itself sends no telemetry, no analytics, no crash reports. Network traffic happens only when:
- You chat. Prompts and project context go to the active provider's backend (xAI for Grok, Anthropic for Claude Code, Google for Gemini), unless you route Grok to a local endpoint or use local models, in which case inference stays on your machine.
- An agent uses a network tool. Web search or fetch, or image generation via xAI Imagine. Automations run with network tools off by default; enabling them is per-rule and explicit.
- You install something. The Grok CLI, FFmpeg, and llama-server are fetched from pinned URLs and SHA-256-verified; installing Claude Code or Gemini runs npm against the public registry.
- You use the browser pane, which loads whatever URL you (or a granted agent) navigate to.
Secrets (the custom-endpoint API key and the local server token) live in the Windows Credential Manager, not in JSON files, and no UI or agent-reachable command can read them back.
The uninstaller removes the app but deliberately leaves your data:
~/.swervebuild/holds chats, projects, automations, recordings, and settings~/.grok/holds the Grok CLI, its auth, memory, and skills
To remove everything: uninstall, delete those two folders, and clear the SwerveBuild entries in Windows Credential Manager (cmdkey /list shows them).
| Path | Purpose |
|---|---|
~/.swervebuild/data.json |
Projects, chats, messages |
~/.swervebuild/automations.json |
Automation definitions |
~/.swervebuild/runs/ |
Per-run transcripts |
~/.swervebuild/attachments/ |
Pasted images, captured stills, recorded clips (budgeted; oldest pruned) |
~/.swervebuild/providers.json |
Provider preferences, with no secrets (those live in Credential Manager) |
~/.grok/ |
Grok CLI (auth, skills, memory, sessions) |
Existing ~/.swervegrok/ data is migrated to ~/.swervebuild/ on first launch.
Windows 10/11 x64, NVIDIA-first for the media features (the recorder falls back to software encoding without NVENC). Personal-first project, best-effort support via GitHub issues. See CHANGELOG.md for what changed per release.
Prerequisites: Node.js 18+, Rust (rustup), and the Tauri Windows prerequisites (WebView2 and MSVC build tools).
git clone https://github.com/Swervebytes/SwerveBuild.git
cd SwerveBuild
npm install
npm run tauri devDo not double-click
target/debug/*.exedirectly. Dev mode needs the Vite server.
| Script | What it does |
|---|---|
npm run tauri dev |
Full desktop app with hot reload |
npm run check |
Type-check Svelte |
npm run test:unit |
Vitest unit tests |
cargo test --workspace (in src-tauri/) |
Rust tests |
npm run test:e2e |
CLI tests (MCP always; ACP if grok and SWERVE_E2E_CWD set) |
npm run smoke:session |
Live drive of the installed app (MCP/CDP) |
npm run release |
Production Windows installer |
The release output lands at src-tauri/target/release/bundle/nsis/Swerve Build_<version>_x64-setup.exe. For a local overinstall: npm run install:local.
Tauri 2 native shell, SvelteKit and Svelte 5 frontend, Rust backend.
src/ SvelteKit UI (routes, components, stores)
src-tauri/src/
acp.rs Multi-chat ACP session pool
jobs.rs Automation runner + scheduler
providers.rs Multi-agent provider registry
provider_auth.rs Agent sign-in surface (ACP authenticate / own flows)
secrets.rs OS keystore seam (Windows Credential Manager)
media_worker.rs Supervised out-of-process capture/encode
local_llm.rs Managed llama-server + VRAM arbitration
live.rs Go-live approval tier (streaming lands in a future release)
store.rs / paths.rs Atomic persistence + data dir
bin/ swervebuild-mcp stdio server, media worker, workflows
crates/swerve-workflows Node-graph workflow engine (Rust + QuickJS)
Contributions welcome. Keep the UI clean and dependencies minimal, and open an issue before large architectural changes. Maintainers: read .github/REPOSITORY_POLICY.md before pushing or cutting a release. Dependabot PRs feed a deliberate upgrade ritual and are never auto-merged.
License: MIT