Run a private AI.
On your own laptop.
One command.
Voice, vision, RAG, web search. No cloud. v0.2.0 out now.
curl -fsSL https://molebieai.com/install.sh | bash
Why I’m Building It
Mission
Build infrastructure and architecture that maximises the performance of LLMs running locally on resource-constrained hardware.
Most local AI tools assume you have a powerful machine and just slap a UI on top. Molebie AI takes the opposite approach — it obsesses over efficiency and optimisation so that even a laptop with 8–16GB of RAM can deliver a genuinely good AI experience. v0.2.0 is out today, shipping voice, vision, RAG, web search, quality gates, and a terminal observability dashboard — all running locally.
The Car Analogy
Imagine you want to build the fastest car, but you can’t afford a bigger engine. So what do you do? You strip out the unnecessary weight, redesign the frame for aerodynamics, and engineer every system around the constraint. That’s the Lotus philosophy — “Simplify, then add lightness.” Molebie AI applies the same thinking to local AI:
Can’t add more RAM?
Use a small fast model for easy tasks and only call the big reasoning model when it matters — like knowing when to use electric vs fuel in an F1 car.
One machine isn’t enough?
Split inference across multiple devices over Tailscale — like having a pit crew ready when you need them.
Heavy databases eat resources?
Use SQLite instead of Postgres — strip out the luxury interior so more power goes to the engine.
Context windows are limited?
Smart RAG, summarization, and memory keep prompts lean and effective — that’s aerodynamics, less drag, more speed.
The Result
Privacy comes for free (everything stays local), but the real goal is making local AI perform its best on hardware that shouldn’t be able to run it well.
What You Get
Three Thinking Modes
Instant answers, chain-of-thought reasoning, or extended deep thinking with an 8K token budget. Opt in to a verification stack — Chain-of-Verification, Grounding Judge, SelfCheckGPT, Self-Consistency voting — when correctness matters more than speed.
Voice Conversation
Wake-word activation, faster-whisper speech-to-text, streaming Kokoro TTS with 12 voices, and speaker verification. Hands-free from the first word.
Image Understanding
Drop, paste, or attach images mid-conversation. Your AI sees and reasons about visual content alongside text.
Document Memory & Vault Sync
PDF, DOCX, TXT, and Markdown with hybrid vector + BM25, cross-encoder rerank, and Anthropic contextual retrieval (+35–49% quality). Folder ingest up to 5,000 files with live progress; hash-diff sync for your Obsidian vault.
Web Search
Self-hosted SearXNG with LLM intent classification, trafilatura page fetch, source trust scoring, and inline citations. Real-time answers without giving up your queries.
Full Data Ownership & Observability
SQLite (WAL + FTS5 + sqlite-vec) on your machine. Multi-user with full isolation. No telemetry by default. Watch every request live from `molebie-ai monitor` — a 7-panel terminal dashboard polling at 2Hz.
See It in Action
CLI tool — one-command install process
Three Thinking Modes
01
Instant
Fast, no chain-of-thought
Instant mode — sub-second responses
02
Thinking
Chain-of-thought reasoning
Thinking mode — 2048 token budget
03
Think Harder
Extended deep reasoning
Think harder — 8192 token budget
Key Features
Voice Conversation
Wake word, faster-whisper STT, streaming Kokoro TTS
Web Search
Self-hosted SearXNG with inline source citations
How It Works
Your Devices— anywhere on your network
PhoneSafari / Chrome
LaptopAny browser
iPad / TabletSafari
Other PCAny browser
Tailscale Tunnelencrypted · 100.x.x.x:3000
A
One Machine
Everything on a single box
Your Machine
Web App:3000
Gateway API:8000
LLM Inference:8080 / :8081
SQLite DB
SearXNG Search:8888
Kokoro TTS:8880
B
Split Across Machines
Coming Soon
Distribute the workload
Machine 1
Web App:3000
Gateway API:8000
SQLite DB
Machine 2 — GPU
LLM Inference:8080 / :8081
Machine 3 — Optional
SearXNG:8888
Kokoro TTS:8880
You decide what runs where based on your hardware.
Got one powerful PC? Run everything there.
Got an old laptop + a GPU tower? Split the workload.
Up and Running in 60 Seconds
01
Install
One-line bash install on macOS, Linux, or Windows WSL. Single-binary CLI, no config files.
curl -fsSL https://molebieai.com/install.sh | bash
02
Auto-Configure
Picks the right backend (MLX on Apple Silicon, Ollama elsewhere) and pulls Qwen 3.5 — 4B for Instant, 9B for Thinking.
Apple Silicon detected → MLX backend → 16GB RAM → Balanced profile
03
Chat
Talk, type, drop in files, sync a vault. Watch it run from `molebie-ai monitor`.
molebie-ai run
Why Self-Host?
Data Privacy
✗
Your data on their servers
✓
Everything stays on your machine, no telemetry
Monthly Cost
✓
Free forever, you own the hardware
Offline Access
✗
Requires internet connection
✓
Works fully offline — search via self-hosted SearXNG
Customization
✗
Limited to what they offer
✓
Swap backends (MLX / Ollama / vLLM / llama.cpp) any time
Observability
✗
Closed dashboard, if any
✓
`molebie-ai monitor` — 7-panel live terminal dashboard
Hallucination Control
✓
Opt-in CoVe, Grounding Judge, SelfCheckGPT, Self-Consistency
System Requirements
Operating SystemmacOS, Linux, or Windows (WSL2)
RAM8 GB min (Qwen 3.5 4B Instant), 16 GB+ recommended (Qwen 3.5 9B Thinking)
GPUApple Silicon (MLX), NVIDIA (vLLM / llama.cpp CUDA), or CPU fallback
Disk Space~10 GB for default Qwen 3.5 4B + 9B
DockerOptional — only for SearXNG search and Kokoro TTS sidecars
Supported Backends
MLXOllamavLLMllama.cppOpenAI-compatible
Frequently Asked Questions
Architecture
User
User
Web App (Next.js)
Chat UIVoiceVisionDocument BrainAuth
Control
Control
Gateway API (FastAPI)
JWT AuthSSE StreamingRAG PipelineWeb SearchVoice
Services
Thinking LLM:8080
Instant LLM:8081
SearXNG:8888
Kokoro TTS:8880
Data
Data
SQLite + Local Storage
AuthUser DataEmbeddingsFilesFTS5