maka: Local-first design, execution log-based agent workspace

Maka is an agent workspace that keeps session settings and execution records locally by default. You can directly choose model connection methods, such as cloud APIs, local models, and compatible gateways, and it provides a desktop app for macOS Apple Silicon, a terminal TUI, and a non-interactive CLI. All interfaces run through the same Runtime Host, ensuring consistent permission management and tool invocation.

The core design philosophy is that 'the log is the runtime.' Model messages, tool calls, tool results, termination facts, and more are recorded in the Runtime Event Log, and the session UI, model context, and recovery features operate as projections of this log. This ensures that recorded evidence is preserved even when Tool Result pruning and LLM Compaction are applied, with only the content visible to the next inference being selectively changed.

It supports local tools such as Read, Write, Edit, Bash, Glob, and Grep, and features execution safeguards including tool schema validation, dynamic availability, permission policies, watchdogs, interruption, and error classification. On the desktop, you can create, archive, search, retry, regenerate, and branch sessions, as well as manage artifact lists and previews, workspace guidelines, and model and permission settings. Local memory, web search, and bot entry points are also supported.

The Eval feature supports declarative multi-arm experiments, scaling across task × repetition × target cells. Attempts within each cell are immutable, and reproducible benchmark experiments are possible through target infrastructure substitution and most-valid selection. It is released under the Apache-2.0 license, with macOS Apple Silicon currently as the official support target and Windows x64 in an unsigned preview stage.