The answer is
fortytwo.
The hard part is knowing
the right question.
fortytwo is a local-first personal-assistant spine for the agents and tools you already use. It wraps an existing harness with memory, a safety gate, channels, approvals, and audit — without replacing the agent you trust.
Two surfaces, one operator.
The npm engine ships the reusable machinery — memory, the gate, channel adapters — as @justfortytwo/* packages wired in as plugins. The scaffolded persona (CLAUDE.md + context/*) is per-user, personal, and gitignored. create-fortytwo / fortytwo is the single CLI over both — and init installs the engine for you.
Seven parts. One spine.
fortytwo is decomposed into independent, composable pieces — each named for the role it plays, each on npm. Bring your own agent; assemble the spine you need. The contract matters more than the adapter.
Conservative by design.
Personal assistants get useful when they can remember, act, and improve. They get dangerous when that’s added without boundaries.
Local-first where it matters
Private memory and recall live on your machine. Markdown for human-readable policy, SQLite for durable state.
Bring your own agent
Claude Code is the first harness, not the boundary. The core stays portable across agents, models, and runtimes.
Conservative autonomy
The assistant may read, draft, and reason internally. External or irreversible actions require approval.
Propose-only learning
It may notice patterns — but never silently promotes them into durable behavior. Propose first, then approve.
Prompt-injection boundaries
Documents, messages, web pages and recalled memory are content, not command authority.
Auditable evolution
Every meaningful change is inspectable — as a file diff, a database record, or an approval decision.