Set up
Runs on your computer. Pick your agent and operating system.
Before you start
- uv (`curl -LsSf https://astral.sh/uv/install.sh | sh`)
{
"mcpServers": {
"leuk": {
"command": "uvx",
"args": [
"leuk"
]
}
}
}same format Cursor, GitHub Copilot CLI, Junie, JetBrains AI Assistant, Augment (Auggie CLI), Factory Droid, Kiro, Devin CLI +12
Values in angle brackets are placeholders. Restart Claude Desktop after editing its config.Claude Desktop docs
README
synced Oct 26, 2026leuk
A persistent AI agent that lives in your terminal.
Multi-provider LLMs · sub-agent orchestration · voice in/out · desktop & browser control · native multimodal (images, audio, video) · shell, files, web & MCP.
Getting Started · Documentation · Configuration · Architecture
Quickstart
uv sync # install
uv run leuk # start the REPL
The default provider is OpenCode Zen (big-pickle) — free, no API key, works
out of the box. Want a different one? Run /auth inside the REPL to set up
Anthropic, OpenAI, Google Gemini, OpenRouter, or a local Ollama/vLLM endpoint.
Everything else is configured from /settings (written to a single
~/.config/leuk/config.json) — no shell environment variables required.
❯ leuk
leuk · zen/big-pickle ctx ~1.2k/200k 1%
❯ summarize today's git log and open a PR with the notes
…
Why leuk
- It remembers. Sessions are SQLite-backed, lazily created, auto-named, and resumable — pick up exactly where you left off.
- It sees and hears. Screenshots, images, audio, and video are sent to the model natively (never base64-pasted as text); voice gives you hands-free speech in and out.
- It does things. Run shell commands, read and patch files, fetch the web, drive a real browser, and control the desktop keyboard & mouse.
- It stays safe. Rule-based review policies, per-tool permissions, persistent approvals, and channel-native approval prompts gate every risky action.
- It scales out. Spawn sub-agents for parallel work, schedule background tasks, and reach it from Telegram, Slack, or Discord.
Highlights
| Providers | Anthropic (incl. OAuth), OpenAI, Google Gemini, OpenRouter, OpenCode Zen, local Ollama/vLLM — context window & capabilities are queried, never hardcoded |
| Multimodal | Images, audio, and video sent through native provider blocks; non-vision models degrade gracefully |
| Tools | Shell, file read/patch-edit, web fetch, sub-agents, memory, plus optional browser and desktop control |
| Extensible | Install SKILL.md agent skills (OpenClaw/ClawHub-compatible) and import MCP connectors from registries — manage with /skills and /mcp |
| Voice | Local or remote STT & TTS with neural VAD; numbers and acronyms are spoken correctly per language |
| Context | Tiered compaction (truncate → mask → summarize → drop) sized to each model's real context window |
| Reach | MCP servers, a background scheduler, and channels (Telegram/Slack/Discord) |
Documentation
The wiki is the authoritative, ground-truth documentation, kept in sync with the code:
- Getting Started — install, extras, first run,
/auth - REPL & Commands · Configuration · Environment Variables
- Architecture · Providers · Context Management
- Tools · Safety · Voice · Multimodal
- Channels · Scheduler · Sessions & Persistence · MCP
- Development
Acknowledgments
leuk's offline voice stack builds on:
- Silero VAD (MIT) — voice-activity detection
- Silero Models — neural text-to-speech (see citation & licensing below)
- OpenAI Whisper (MIT) — speech-to-text
Citations
@misc{Silero Models,
author = {Silero Team},
title = {Silero Models: pre-trained text-to-speech models made embarrassingly simple},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/snakers4/silero-models}},
email = {hello@silero.ai}
}
Silero model licenses & commercial use
leuk is AGPL-3.0. The Silero TTS models are downloaded at runtime (not bundled in this repo), so they do not alter leuk's license — but their own licenses constrain how you may use the synthesized output:
| Component | License | Commercial use |
|---|---|---|
| Silero VAD | MIT | ✅ yes |
Silero TTS — CIS (v5_cis_base: ru, uk, kk, tt, uz, ky, ba, xal) | MIT | ✅ yes |
Silero TTS — other (v3_en, v3_de, v3_es, v3_fr, v4_indic) | CC BY-NC | ❌ non-commercial only |
So Russian and CIS-language TTS is free for commercial use; English, German, Spanish, French, and Indic TTS via Silero is non-commercial only. For commercial use of those languages, switch to the OpenAI TTS backend (or a permissively-licensed engine such as Piper). See Voice for details.