An AI kernel that hijacks Bash to control your native apps

What it does

Yome Agent is an open-source runtime that wires an LLM into your local apps through a single Bash abstraction. It lets the model invoke native software—PowerPoint, Excel, Calendar—via AppleScript, Win32, or DBus by intercepting shell commands and routing them to installed skills. The broader vision is an always-on daemon that watches your filesystem and messages, but today the kernel, skill sandbox, and thread history are the solid ground.

The interesting bit

The clever part is the three-layer documentation system. Skill maintainers author an L1 index—just three lines that sit in the system prompt—so the model knows when to fire the tool, what side effects to expect, and what to type first. If the task gets complex, the kernel escalates to L2 signatures or L3 cookbooks only when needed, keeping the token budget tight. Batched PowerPoint tasks run roughly six times faster than sequential ones when merged into a single osascript process, at least on an M1 Mac.

Key highlights

• One Bash tool handles real shell commands, skill verbs, and piped composition alike; the kernel inspects the first token to decide where to route it.
• Native skills declare sandboxed capabilities—fs:write, applescript, network—and the kernel returns a hard error if the model tries an ungranted one.
• The L1/L2/L3 doc layers keep system prompts small (about 60 tokens per skill) while still giving the model full cookbooks for non-trivial tasks.
• @yome/ppt is the only stable skill so far, with 16 actions and batch merging; Excel, Calendar, Reminders, and Mail remain beta or alpha.
• Thread history and live compaction work today; the always-on daemon and event-driven “Oncall” features are still on the next branch.

Caveats

• The always-on daemon and event-driven “Oncall” features are listed but not yet available; the README notes they are “next-up on the next branch.”
• Outside of PowerPoint, the native skills are explicitly tagged beta or alpha.
• Even the stable PowerPoint skill can trigger an OS permission dialog on first save to the Desktop, so fully headless automation is not guaranteed.

Verdict

Developers who want their LLM to actually click around in local apps rather than just suggest code should look here—especially if you live in a shell. If you need a fully autonomous, always-on digital assistant today, wait for the next branch to land or look elsewhere.