Claude, GPT, Gemini, Grok, Kimi, Qwen, GLM, MiniMax, Hunyuan — plus local models. One quiet desktop app for every AI you want to talk to. No lock-in, no subscription.
⌘K focus composer · ⌘N new session · Esc close
Everything here ships today. Free and open source.
| No subscription | $0 app fee. You pay only the providers you already have keys for — or pay nothing when you run local models. |
| 10 providers | Anthropic (Claude 4.7/4.6/Haiku 4.5), OpenAI (GPT-5.4/Thinking/Pro/5.3), Google (Gemini 3.1 Pro/Flash), xAI (Grok 4.20), Kimi (kimi.ai + kimi.com, K2.6), Alibaba Qwen 3.6 Max, Zhipu GLM-5, MiniMax M2.7, Tencent Hunyuan Hy3 Preview. One dropdown. |
| Local via Ollama | Fully offline when you want. In-app catalog with one-click pulls for Qwen 3.6, Gemma 4, Llama 3.2/3.3, DeepSeek R1. Installed models appear alongside cloud ones in the same picker. |
| Crew — multi-model in parallel | Send one prompt to 2–4 models at once. Each replies in its own lane. Pick the best, or have them debate each other. Useful when one model's bias isn't enough. |
| Skills — cron for AI | Schedule Labaik to do things on its own. "Summarize HN at 8am." "Draft my standup from git log at 6pm." "Ping prod every 15 minutes." Results stream to a dedicated session. |
| Built-in browser | The agent can open URLs, read pages, click buttons, fill forms, and take screenshots — all inside the app. Research, QA, and web-scraping without leaving chat. |
| Rich content display | Real markdown, fenced code with copy buttons, unified diffs, charts, Mermaid, SVG, file previews, artifacts. Whatever the model returns, the app renders it right. |
| Persistent memory + profile | "Remember this" button on any message. Profile = always-on facts about you. Memory = workspace-scoped, searchable with embeddings. Survives every crash (atomic file writes to ~/.labaik/). |
| Tool calling + MCP | Six built-in workspace tools (read_file, write_file, list_directory, run_command, open_in_browser, start_dev_server) plus full MCP server support for anything else. |
| Privacy by construction | Developer-ID signed with hardened runtime. No Labaik backend. API keys in ~/.labaik/credentials.json at mode 0600. Prompts go machine → provider directly. Zero telemetry sent to anyone. |
Six capability pillars. None of them bolted on — each one is load-bearing in the product.
Anthropic (Claude 4.7 / 4.6 / Haiku 4.5), OpenAI (GPT-5.4 + Thinking/Pro + 5.3), Google (Gemini 3.1 Pro / 3 Flash), xAI (Grok 4.20), Kimi (kimi.ai + kimi.com, K2.6), Qwen (3.6 Max Preview), GLM-5, MiniMax M2.7, Tencent Hunyuan Hy3. Your keys, in ~/.labaik/credentials.json at mode 0600.
In-app Ollama catalog with one-click pulls. Qwen 3.6, Gemma 4 (E2B/E4B/26B/31B), Llama 3.2/3.3, DeepSeek R1. Installed models land in the same dropdown as cloud ones.
Claude + GPT + Gemini arguing until they get it right. Send one prompt to 2–4 models in parallel, pick the best reply, or have them debate each other. No other desktop AI does this.
Schedule Labaik to do things. "Summarize Hacker News at 8am." "Draft my standup from git log at 6pm." "Ping prod every 15 min." Runs in the background, results stream to a dedicated session.
Profile (always-on facts) + scoped memory with embeddings. Workspace-aware. "Remember this" button on any message. Persisted to ~/.labaik/*.json — survives every crash.
The agent can open any URL, read pages, fill forms, click buttons, and take screenshots — all from inside the app. Research without leaving chat.
Six workspace tools: read_file, write_file, list_directory, run_command, open_in_browser, start_dev_server. Plus MCP server support for anything else.
No Labaik backend. Prompts go machine → provider you picked. Telemetry stays on disk. Renderer ↔ Main isolated via contextBridge. No nodeIntegration. Developer-ID signed + hardened runtime.
Real markdown, fenced code with copy buttons, hover-to-copy, smart auto-scroll, voice input, drag-and-drop attachments (PDF, DOCX, XLSX, images), keyboard shortcuts, Cmd+F search across all sessions.
Labaik is provider-agnostic. Switch in a dropdown. Tool calling is gated per model, so tiny models never get tools they can't format.
API keys at ~/.labaik/credentials.json, mode 0600. Memory, profile and session history at ~/.labaik/*.json. Nothing leaves your machine except outbound requests to the provider you picked.
There is no Labaik server to talk to. Prompts go straight from your machine to whichever LLM provider you selected. The OODA event log stays on disk at ~/.labaik/events.ndjson. Zero telemetry sent to anyone.
Developer-ID signed + hardened runtime. Renderer ↔ Main isolated via contextBridge. No nodeIntegration. Markdown renderer escapes HTML before transforms. Health Space carries a not-medical-advice reminder.
Developer-ID signed · hardened runtime · builds attached to every GitHub release. Windows and Linux builds coming soon.
M1 / M2 / M3 / M4 · 129 MB
x86-64 · 136 MB
NSIS · AppImage · deb
git clone [email protected]:alsayadi/alaude-desktop.git
cd alaude-desktop
bun install
bun startYes — at least one. Anthropic for Claude, OpenAI for GPT-5, Google for Gemini, etc. Create a key at the provider's website (e.g. console.anthropic.com / platform.openai.com), paste it into Labaik's Keys modal, done. You can connect multiple providers and switch in a dropdown. Or skip keys entirely and run Ollama for fully-local, fully-offline.
Because Cursor charges $20/mo on top of what you already pay providers. You shouldn't pay twice. Labaik is just a desktop UI over the APIs you already have — no middleman, no markup, no data-mining. If you never open it, it costs zero.
Yes — Labaik talks to them directly via their official OpenAI-compatible APIs. Create a key at the relevant platform (platform.kimi.ai / kimi.com, dashscope.aliyun.com, open.bigmodel.cn, platform.minimax.io, hunyuan.tencent.com), paste it in. No VPN needed from outside China if you use the kimi.ai endpoint.
No. There is no Labaik backend. Prompts go straight from your machine to whichever LLM provider you selected. The OODA event log stays on disk at ~/.labaik/events.ndjson.
Yes — install Ollama, pull a model through the in-app catalog, and you're done. Tool calling works on capable local models (Qwen 3.6, Gemma 4 E4B+, Llama 3.3, DeepSeek R1). Tiny models have tool calling disabled on purpose — they can't format it reliably.
Multi-model mode. Send one prompt and 2–4 models reply in parallel (e.g. Claude + GPT-5 + Gemini), each in their own lane. Pick the best reply or have them debate each other. Useful for consequential decisions where one model's bias isn't enough.
Cron for AI. Schedule Labaik to run a prompt on a recurring schedule — "summarize HN at 8am," "draft my standup from git log at 6pm," "check prod status every 15 minutes." Results stream to a dedicated session you can scroll back through. Runs in-process (no cloud daemon), so your machine has to be awake.
Two layers. Profile = always-on facts about you (name, role, preferences) injected into every turn. Memory = scoped, searchable with embeddings, filtered by workspace so your work stuff doesn't bleed into your personal stuff. Both persisted to ~/.labaik/*.json with atomic writes — survives every crash.
In ~/.labaik/credentials.json with file mode 0600. Only your user can read them. They never leave the machine except as outbound requests to the provider you picked.
Open any URL, read the rendered page, click links, fill forms, type into inputs, take screenshots, and scroll. Useful for research ("summarize this Hacker News thread and save the links I should read later"), QA (navigate a deploy preview and report layout issues), or dull web-scraping tasks. The browser is containerised inside the app, not a popup that steals focus.
Neither. Different surface (desktop, not IDE), different audience (everyone, not just coders), different codebase. Labaik takes the lesson of those products — good tool loops are mostly prompt engineering — but is written from scratch, MIT licensed, source-available.
MIT. Source at github.com/alsayadi/alaude-desktop.