M1K3 is local AI that runs entirely on your Mac. Local models, live voice, a knowledge graph that remembers everything — no account, no telemetry.
Everything below runs inside one sandboxed app on Apple Silicon. Pull the ethernet cable — M1K3 doesn't notice.
Three brains: Mini answers instantly, Lil fronts the conversation, and Big is delegated the deep work. Switch in Settings any time — which models, and how they score.
mlx-swift · afm · qwen3 · gemma-4On-device speech recognition in; Kokoro neural TTS out — pure MLX, streamed sentence by sentence, with karaoke captions. M1K3 makes its own end-of-turn call, and ending on “please” hands it the floor.
kokoro · mlx · whisperkit · say pleaseImport documents and M1K3 builds a private knowledge graph with semantic search. Ask questions; get answers grounded in your own files.
rag · bge embeddings · ftsConsent-gated call recording with on-device transcription, speaker-grouped summaries, and encrypted persistence. Your calls become searchable knowledge.
encrypted · keychain · consent-firstA built-in ReAct agent with real tools — web search, page reading, your knowledge base — behind a privacy toggle. Web off means web off.
local agent · tool calls · privacy toggleM1K3 exposes your knowledge, memory, and voice to Claude and other agents over the Model Context Protocol — with an async job API for the slow, deep asks. Your second brain, available to your other AIs — locally.
mcp · 18 tools · 127.0.0.1:4242A temporal graph of dated facts. Correct M1K3 and the dream cycle supersedes the stale fact instead of silently losing either version — history stays visible.
temporal graph · dream cycle · supersede, don't deleteEvery two hours, a short narrative pulse about your machine and your day — facts from code, never confabulated, quiet hours respected. Off by default; never enters your chat.
ambient · quiet hours · off by defaultServe your Mac's brain to your own devices over the LAN: Bonjour discovery, one-time QR pairing, TLS-PSK encryption, and a scoped read-only tool set. Everything stays inside your walls.
tls-psk · bonjour · qr pairing · off by defaultNot an Electron wrapper. Not a webview. A SwiftUI app with Liquid Glass, RealityKit avatars, and Metal-accelerated inference — built for macOS Tahoe and nothing else.
Not by default. M1K3 runs on your Mac — inference, voice, embeddings, and storage are all on-device. Turn on Private Cloud Compute in Settings and a message you choose to send — nothing else — goes to Apple's Private Cloud Compute; every PCC reply is labelled in the chat, and Apple's own guarantees for that service are quoted on our privacy page. Even the thumbs-up/down feedback you leave on answers stays local — it's yours, for tuning, and no visiting agent can read or write it. Web search is on by default so answers can be current, with one switch in Settings to turn it off.
Three brains, all local: Mini for instant answers, Lil — the fast brain that fronts the conversation — and Big, delegated the deep work. Switch in Settings any time; the models behind each name, and their scores, live on /brains.
Only if you ask it to. The Heartbeat is an ambient pulse — every two hours, a short narrative about your machine and your day, composed from facts the code gathered, respecting quiet hours. Off by default, one-tap Clear, never enters your chat transcript. There's also a native screensaver — the pixel M in ambient rain — installable from Settings.
That's Brain at Home: your Mac serves its brain to your own devices over the local network — Bonjour discovery, one-time QR pairing, TLS-PSK encryption, and a scoped read-only tool set for paired devices. Off by default; nothing ever crosses the internet.
Apple Silicon Macs on macOS 26 Tahoe. M1K3 is a native SwiftUI app with Metal-accelerated inference — not an Electron wrapper.
Ollama is a local model server; ChatGPT is a cloud service. M1K3 is a complete on-device AI app: local models plus live voice, a knowledge graph over your documents, encrypted call memory, an agent with tools, and an MCP server — one sandboxed native app.
Yes. M1K3 ships an MCP server with 18 tools — search_knowledge, ask_m1k3, memory recall, documents, even speak — over local HTTP at 127.0.0.1:4242. Long asks get an async job API. Claude and other MCP clients query your local knowledge base, locally.
Free for humans, forever — made by Round Tower in Ireland. No account, no subscription, no telemetry. It is source-available, not open source: the whole repository is public under the Functional Source License: read every line, build it, run it for yourself or inside your organisation; the one thing it withholds is selling M1K3 or a copy of it to others, and each version becomes Apache-2.0 two years on. Download the signed & notarized Mac app, join the TestFlight beta, or build from source today. If M1K3 earns a place on your Mac, there's an optional GitHub Sponsors page — it funds development, never gates features.
Organisations: M1K3 for Teams puts the same brains, an institutional memory, and a persona tuned to your people on your own Apple Silicon hardware, under your own domain. An administrator can lock the Private Cloud Compute switch off fleet-wide, so no conversation goes to anyone's cloud. Licensed separately; write to hello@round-tower.ie.
M1K3 for Mac is ready to download. Free for humans, forever — run it on your own Mac, or build it from source. Two minutes, no account: get the app, say hello to the brain already on your Mac, and — if you want — connect your coding agent with one command.
Free, and staying that way. If it earns a place on your Mac, you can sponsor its development — optional, never nagged.