What it does Moltis is a single Rust binary that runs a persistent AI agent on your own machine — Mac Mini, Raspberry Pi, or server. It connects to multiple LLM providers (OpenAI Codex, GitHub Copilot, local models), remembers context across sessions, and meets you wherever you already chat: Telegram, Signal, WhatsApp, Discord, Teams, Matrix, even Nostr. Voice I/O, browser automation, scheduled tasks, and MCP tool servers are built in, not bolted on.

The interesting bit The security posture is unusually paranoid in a good way: API keys never leave your machine, every tool call runs in a sandboxed container, and the core agent loop is safe Rust with unsafe code quarantined to FFI boundaries. The 59-crate workspace is deliberately over-documented — ~270K lines you can actually audit, with artifact attestations and Sigstore signing for releases. It’s the anti-SaaS: no plugin marketplace to get supply-chain attacked through.

Key highlights

• One binary, no Node.js/npm runtime; Docker or bare metal
• 8 TTS + 7 STT voice providers, mobile PWA with push
• SQLite + vector + full-text memory with cross-session recall
• Automatic checkpoints before any file edit or memory mutation
• MCP servers via stdio and HTTP/SSE, plus OpenClaw skill import
• --features lightweight strips down for constrained devices

Caveats

• The README is confident but light on actual performance numbers or latency benchmarks
• 270K LoC is substantial; the “small unsafe surface” claim is well-structured but you’ll need to trust the crate boundary discipline
• Rust 1.91+ required; no mention of Windows support in the install paths shown

Verdict Grab this if you want a full-featured AI assistant but can’t stomach cloud-hosted agents with your code and keys. Skip it if you’re looking for a quick managed solution or need heavy Windows-first tooling.