Skip to content

README.md

Latest commit

 

History

History
373 lines (248 loc) · 14.9 KB

README.md

File metadata and controls

373 lines (248 loc) · 14.9 KB
Raw
warpdrv

warpdrv

Local LLM Server Manager + LlaMA.cpp + Chat + More

😎 Built for Local AI Master Race! 😎

License: AGPL v3 Release Build Stars Issues Platform Discord Reddit


Screencast.From.2026-06-25.17-44-11.mp4
warpdrv hero screenshot

Alpha release - expect things to be broken. This project is in active development.

What is warpdrv?

warpdrv is a desktop toolkit for using local LLMs.

  • ⭐ Use Local LLMs ⭐ Chat with local models, or integrate them with your favorite AI tools.
  • ⭐ Rich Chat UI ⭐ A polished chat interface supporting Elicitation, Graphs, Syntax Highlighting and more.
  • ⭐ Voice ⭐ Use Kokoro Text-To-Speech, Whisper Dictation, or combine them in Voice-Chat mode.
  • ⭐ Built-in MCP ⭐ Use tools that come with the built-in MCP server. Also supports external mcp.json.
  • ⭐ RAG ⭐ Build projects using Workspace and Embedding using local models.
  • ⭐ Power Your Workflows ⭐ Features such as Annotations, Guardrails, Compaction to drive your daily tasks.
  • ⭐ Multi-GPU Multi-Vendor ⭐ Use any CUDA/ROCm/Vulkan hardware.
  • ⭐ Build your own LLAMA ⭐ Supports custom-built LLAMA.cpp.
  • ⭐ Open Source ⭐ No hidden code. No analytics.
Screencast.From.2026-06-25.17-51-23.mp4

Table of Contents

Screenshots

Chat interface Model browser
Llama backends Hub
More screenshots
Onboarding welcome screen Launch server dialog
Llama build Recipe MCP tool calls

Features

Server management. Launch llama-server instances with full parameter control, including cache checkpointing, speculative decoding, n-gram speculation, multimodal projections, and any custom flag you need. Per-model parameter overrides let you save the right settings for each model so launch is one click.

Proxy. An OpenAI-compatible endpoint proxy lets any chat app talk to your running servers. Server aliases route requests to the right backend, with optional auth and a built-in web server for the UI, so you can access host and warpdrv anywhere.

Backends. Register one or more llama-server builds — stock, custom-compiled, ROCm, CUDA, whatever — and group them for quick swapping. The Recipe Engine compiles fresh builds on demand using shared bash recipes.

Models. Browse and download from Hugging Face Hub directly inside warpdrv, with full download management. Local models are scanned, parsed, and organised into folders.

Chat. A built-in chat UI with threads, folders, and full sampling configuration. MCP server integration handles tool calling with per-tool permission prompts.

Slash Commands. Warpdrv comes with built-in slash-commands to perform tasks such as context compaction, using to-do lists, security and compliance guardrails using parallel adversarial local models.

Voice You can launch Whisper servers for STT and use KokoroJS for TTS. The are combined together in a powerful voice chat mode which makes every conversation feel like a phone call. Comes with 7 different voices.

Install

Linux (recommended)

Download the latest .deb or .AppImage from the releases page.

.deb (Debian, Ubuntu, Mint):

sudo dpkg -i warpdrv_*.deb

.AppImage (any distro):

chmod +x warpdrv-*.AppImage
./warpdrv-*.AppImage

Windows

Download the latest .msi installer from the releases page.

Note: Recipes feature does not work on natively Windows as it relies on Bash scripts.

macOS

No prebuilt build yet. Build from source — see For Developers. Untested on Apple Silicon; PRs welcome.

First Run

On first launch, warpdrv shows an onboarding welcome screen that walks you through:

  1. Adding a folder where your GGUF models live (or where to download them)
  2. Optional: registering a llama.cpp build
  3. A short slideshow of next steps (launch a server, start chatting, etc.)

Config and data are stored at ~/.config/warpcore/. Survives upgrades — config is preserved across reinstalls.

Updating

warpdrv checks for updates on startup and shows a banner if a new version is available. Click the banner to open the releases page, download the new version, install it. Your data is preserved.

Quick Start

  1. Install warpdrv — see Install
  2. Onboarding — pick a models folder, optionally add a llama.cpp build
  3. Scan models — warpdrv parses every GGUF in your folder. Alternately navigate to the Hub page to download a model.
  4. Add a backend — point warpdrv at a llama-server binary; it auto-detects devices
  5. Launch a server — pick a model, set GPU layers + context, hit Launch
  6. Open Chat — pick the running server, start a thread, test the model

How-To Guides

Hardware Compatibility

Tested Configurations

warpdrv works with any standard llama-server binary, so compatibility tracks llama.cpp's own support matrix.

Backends

Backend Status
CUDA (NVIDIA) Supported
ROCm (AMD) Supported
Vulkan (any GPU) Supported
CPU only Supported

Notes

  • Bring your own llama-server binary built against your hardware. The Recipe Engine can help compile one.
  • Speculative decoding may not work with sliding-window-attention models.
  • For GPU-specific build flags and runtime quirks, see the llama.cpp documentation for your target.

FAQ

Is this a daily-driver chat app? Depends. warpdrv is a workshop bench for testing models and llama.cpp builds. For polished daily chat, or coding, warpdrv's proxy server integrates seamless with your existing tools and workflows, and provides a customizable router using user-defined aliases.

How does this differ from LM Studio / Ollama / Jan? Many server management apps bundle a fixed llama.cpp version and limits which models / quants work. warpdrv lets you bring your own llama.cpp builds — including bleeding-edge forks — and run them with full parameter control. Built for tinkerers, and also for end-users wanting a one-click chat app.

Do I need to compile llama.cpp myself? No, but you can. warpdrv works with any standard llama-server binary. The Recipe Engine helps if you want to compile your own. Note: warpdrv does not ship with a llama binary, you have download one from the official github repo LlaMa.cpp Releases.

Where is my data stored? Linux: ~/.config/warpcore/ — chat database, settings, MCP config, recipes. Models stay wherever you put them; warpdrv only indexes them. Your data never leaves your computer.

Can I use warpdrv commercially? warpdrv is licensed under AGPL-3.0. If you offer it as a network service, you must publish your modifications under AGPL. For commercial licensing without AGPL obligations, join the Discord and PM the mods.

Why AGPL? To keep the project genuinely open: derivatives stay open, including SaaS forks.

Does warpdrv send my data anywhere? No telemetry, no analytics, no remote calls — except the update-check ping to fetch release.json from the GitHub repo on startup.

Why is X feature broken? Alpha software. File an issue with reproduction steps. Better yet, send a PR :)

Who made this app? A combination of AIs and human effort by one software developer. AI used were local Qwen 3.5 27b, Qwen 3.5 27b, Qwen 3.6 35b, Claude Opus 4.6 for planning and skeletons. This app itself was used to host the local LLMs during development.

How long did this take to build Approx. One week for MVP. One month for initial public release.

Roadmap

  • Short-term
    • Stabilise critical features (server stop, log parsing)
    • Windows prebuilt installer
  • Mid-term
    • macOS prebuilt build (Apple Silicon)
    • Voice dictation in chat
  • Long-term
    • Richer chat interface.

User feedback and feature requests are very welcome — drop them in Discord, Reddit, or GitHub Issues.

For Developers

Architecture

warpdrv is a Tauri desktop app wrapping a Node.js server and a React frontend. The Tauri shell spawns the Node server as a sidecar binary on launch, monitors its health, and restarts it on crash. The React app talks to the server over HTTP + SSE.

Tech Stack

  • Desktop shell — Tauri 2 (Rust)
  • Frontend — React 19, Chakra UI v3, Vite, Zustand, assistant-ui
  • Server — Node 24, Express 5, better-sqlite3, better-sse
  • Bundling — esbuild + @yao-pkg/pkg (server binary), Vite (frontend)
  • Shared types — TypeScript-only @warpcore/shared package

Monorepo Structure

packages/
  shared/   @warpcore/shared   — Types, enums, utilities. No runtime deps.
  app/      @warpcore/app      — React frontend.
  server/   @warpcore/server   — Express + SQLite. Process management, GGUF parsing, recipes.
  bridge/   @warpcore/bridge   — Composable chat engine (extracted, used internally).
  desktop/                     — Tauri shell + release scripts.

Build From Source

Prerequisites:

  • Node 24+
  • Rust + Cargo (for Tauri)
  • Linux: standard Tauri dependencies — see Tauri prerequisites

Steps:

git clone https://github.com/mikjee/warpdrv.git
cd warpdrv
npm install

Run in dev mode:

The recommended way is via VSCode. Open the repo, go to the Run and Debug panel, pick the warpdrv-all launch config, and hit play. All packages launch in a single integrated terminal.

If you don't use VSCode, run the same thing manually:

npm run dev

Release build:

./release.sh                    # Linux .deb (default)
./release.sh deb appimage       # both .deb and .AppImage
./release.sh appimage           # AppImage only

Bundle formats supported by Tauri: deb, appimage, rpm, dmg, msi, nsis, updater. Pass any combination to release.sh.

Artifacts land in packages/desktop/target/release/bundle/.

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup, conventions, and PR rules.

Community & Support

Acknowledgements

Thanks to the following projects and their developers/maintainers for enabling developers to create wonderful Open-Source projects:

Sponsor

warpdrv is built and maintained by me, a solo dev. Consider supporting my open-source work - it helps in funding the hardware and fuels the late-night coding sessions 🤓

Sponsor

License

warpdrv is licensed under the GNU Affero General Public License v3.0 — see LICENSE.

In plain English:

  • Free for personal, hobbyist, and internal commercial use
  • Modifications must be shared under AGPL if you distribute or run as a network service
  • Original copyright notices must be preserved
  • No royalties owed; no warranty provided

For commercial licensing without AGPL obligations, join the Discord and PM the mods.