M1K3
app.m1k3v1.0.0更新於 Oct 3, 2026
Private memory, documents, voice and a local LLM for your agents, served by a Mac app on-device.
概覽
一款 Mac 應用程式,透過本機 MCP 端點向其他代理提供裝置端記憶、文件、語音與本機 AI 助理能力。
- 功能
- M1K3 是一款 macOS 應用程式,在回送位址上透過 HTTP 執行自己的本機 MCP 伺服器。它提供以個人知識圖譜與 RAG 為基礎的知識檢索、文件存取、語音功能,以及用來詢問常駐裝置端 AI 的 ask_m1k3 工具。推論透過 MLX 在本機執行,並可選擇下載更大的模型。
- 適用情境
- 當你希望代理使用私密的裝置端知識庫、文件與本機模型,而非雲端服務時,它很有用。適合想把記憶與語音能力保留在自己電腦上的 Mac 使用者。
- 執行需求
- Apple Silicon 上的 macOS 26 (Tahoe)、Xcode 26.x 與 Swift 6.2 工具鏈,以及 XcodeGen(可選 xcbeautify)。應用程式需從原始碼建置並使用個人開發團隊簽署。MCP 伺服器在本機回送 HTTP 位址上提供服務;模型需從 Hugging Face 下載一次,網頁搜尋預設為開啟。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 M1K3,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
M1K3 for macOS — build from source
The flagship surface: a Mac-native SwiftUI app — on-device MLX inference, live voice (Kokoro TTS + WhisperKit STT), a personal knowledge graph with RAG, an embedded agent, a 3D companion, and a local MCP server other agents can call.
Private by design. M1K3 runs on your Mac. Models download once from Hugging Face; after that the only network use is web search, which is on by default so answers can be current, with one switch in Settings to turn it off. (An opt-in Apple Private Cloud Compute rung arrives in a later release; the 1.0 build has no cloud path.)
Requirements
- macOS 26 (Tahoe) on Apple Silicon — hard requirement: the app uses Liquid Glass, on-device Foundation Models, and Metal via MLX.
- Xcode 26.x (Swift 6.2 toolchain).
- XcodeGen and (optional, nicer logs)
xcbeautify:
brew install xcodegen xcbeautify - Disk: nothing to download for the default brain (Mini runs on Apple's on-device models). Allow ~2–7 GB extra only if you opt into the bigger Lil/Big brains.
Build & run
Or from the command line:
First launch is one screen: say hello (a name is optional) and you're talking — M1K3 starts on Mini, Apple's built-in on-device model, so there's nothing to download. Sharper local brains (Lil / Big) are one click away in Settings, or via the in-chat offer after your first answer; weights download only when you ask.
Signing note: run with your personal development team. Building with
CODE_SIGNING_ALLOWED=NOcompiles fine (CI does it) but the app then runs without its sandbox container and will read/write the wrong data locations.
Tests
The business logic lives in SwiftPM packages and tests without Xcode:
Tests use swift-testing (import Testing, @Test), not XCTest.
Heavy MLX/Metal integration tests are env-gated off by default —
M1K3_MLX_INTEGRATION=1 swift test opts in (downloads a real model and
generates; Apple Silicon only).
Where things are
MCP server
The running app serves MCP over HTTP at http://127.0.0.1:4242/mcp —
knowledge search, documents, voice, and ask_m1k3 (ask the resident AI).
See docs/MCP_SETUP.md to wire it into Claude Code or
any MCP client.
Contributing
See CONTRIBUTING.md at the repo root. Found a bug?
In the app: Settings → Report an issue… generates a redacted diagnostic
you can attach to a GitHub issue.
來源:macos/README.md,提交 898ec88
工具
0版本歷史
1- v1.0.0最新Oct 3, 2026


