LePika

One command → local AI chat in your browser.
(luh-PEE-ka)

Your gaming GPU is already an AI machine. LePika self-hosts LLMs on the hardware you already own — Mac (Metal), Linux and Windows (NVIDIA) — with zero configuration. No Docker, no API keys, no monthly bill, and nothing you type ever leaves the box.

Named after Pika, the cat who supervises this project. Like any cat, Pika has everything worth having at home and no use for the cloud. Same idea, for your AI.

Install

Mac / Linux

curl -fsSL https://raw.githubusercontent.com/MikaelSabuhi/lepika/main/install.sh | sh

Windows (PowerShell)

irm https://raw.githubusercontent.com/MikaelSabuhi/lepika/main/install.ps1 | iex

Already have uv?

uv tool install lepika

Then run lepika. It detects your OS, GPU, and RAM, offers the models that actually fit your machine, installs Ollama + OpenWebUI, pulls your pick, and opens the chat UI at http://localhost:3000.

LePika installs into ~/.local/bin — if a new terminal can't find the lepika command, add that directory to your PATH or restart your shell.

Prefer to read the source first, or hack on it? git clone https://github.com/MikaelSabuhi/lepika && cd lepika && uv run lepika — git and uv are the only prerequisites; uv fetches its own Python.

What it does

• Native GPU, no containers — Apple Metal on macOS, CUDA on Linux/Windows. Ollama arrives via Homebrew, the official install script, or winget; OpenWebUI via uv tool. CPU-only machines work too (LePika warns you they'll be slow).
• Models that fit — a curated list filtered to your RAM, so there's no guessing whether a 27B fits in 16 GB.
• Any model, one field — Ollama tags, Hugging Face GGUF builds, or full-weight safetensors repos. Model guide →
• Private by default — Express listens on 127.0.0.1, full stop. Sharing the stack with your network is a Server-mode feature (lepika expose).
• Two modes — ⚡ Express (default): everything native, no Docker. 🐳 Server: the same thing as one readable docker compose stack, for the box under the desk. Server guide →

LePika reuses an Ollama you already have instead of installing a second copy.

Everyday commands

Global flags: --version, --mode express|server (the wizard's, not a per-command switch), --dry-run, and --help on every command.

All state lives in one place: ~/.lepika — config, logs, and (in Express mode) OpenWebUI's own chats and uploads. That is what makes lepika update keep your chats. Point LEPIKA_HOME somewhere else if you prefer.

Why not just use Ollama directly?

You absolutely can. LePika drives Ollama and OpenWebUI; it doesn't replace them. What it saves you:

If you'd rather run the raw tools, their docs are excellent and LePika gets out of your way. Curious what it's doing on your behalf? It's a few files of plain Python in src/lepika/: three dependencies, no magic.

Requirements

• Disk for the models: ~0.5 GB for a tiny one, ~5 GB for a good all-rounder, 40 GB+ for the flagships.
• 8 GB RAM recommended. Less works: LePika steers you to smaller models.
• A GPU is optional. CPU-only machines run everything, slowly, and LePika says so up front.
• macOS: Homebrew (without it, LePika points you at the Ollama.app download). Windows: winget, which ships with Windows 10/11. Full-weight imports have their own fine print — see the model guide.

Documentation

• Model guide — every model-ref shape, full-weight imports, gated repos
• Server mode — the compose stack, using a GPU box remotely, security
• Architecture — design rules and how it's built
• Changelog — what changed in each release

Contributing

Issues and pull requests are welcome — including an issue for what you'd want next. CONTRIBUTING.md has the short version of how a change gets in. Found a security problem? SECURITY.md says how to report it privately.

uv sync --dev && uv run pre-commit install && uv run pytest -q

License

MIT © Mikael Sabuhi