Maka

Model messages, tool calls, tool results, permission decisions and how a turn ended are written to an append-only log, and both the interface and the next model call are views onto that log rather than the only copy of it. Trimming context therefore drops old tool output from what the model sees without deleting the evidence, and an interrupted turn can be resumed after a crash. Sessions, settings and run records stay on the machine by default, and the model comes from wherever you point it: a cloud API, a local model, or a compatible gateway.

Desktop, terminal and the evaluation harness all execute through one Runtime Host. The built-in tools are Read, Write, Edit, Bash, Glob and Grep; anything leaving the sandbox needs approval, runs can be aborted, and failures are classified rather than dumped. Sessions can be archived, searched, renamed, retried, regenerated or branched from a single turn. The project is incubating at the Apache Software Foundation under Apache 2.0, with releases for Apple Silicon, an unsigned Windows preview, and no Linux build yet.