Command line and MCP assistants

kabacus-cli is the app without a window. It calls the same libraries, so a script or an assistant gets the same validation, revisions and conflicts as a person.

Subcommands

info, new, report, import, export, history, run, commands and mcp. Reading subcommands take --json. --help prints to stdout and exits 0; usage errors exit 2 before any work; failures exit 1.

Scripts

run executes a script, one step per command. A step can name itself with “as” and later steps use its result. A script is all-or-nothing and saved as one revision. Writes are dry runs unless you pass --save. kabacus-cli commands lists every action.

MCP

• Headless: kabacus-cli mcp --books FILE serves MCP over stdio on the books you name. --read-only lists only reading tools. It never reaches the network.
• Attached: kabacus-cli mcp --attach PORT drives a running app through its control channel so you watch every change. Turn the channel on in Settings → Automation. It listens on loopback only.
• Tools are the same everywhere, dry run by default; rewrite tools are listed only when usable.
• A revision made through MCP records your user and the agent’s name, for example “jordan.lee via Claude (agent)”, and Activity marks it.

Review automation like any other write

Read-only MCP exposes only read tools. Write tools use dry runs by default, but a caller with permission can request a real write directly: there is no confirmation token that guarantees a human reviewed it. Give assistants only the books and capabilities they need and inspect their revisions in Activity.

Attached automation is opt-in and binds to loopback. While an agent is attached the sidebar identifies it and changed rows are briefly outlined. A browser tab does not expose that channel.

• Command-line reference
• Amounts, balancing and validation