A while back I wrote up simple tools as an umbrella for a few Rust libraries I kept reaching for — a TUI layer, a networking layer, a capability-scoped shell. It’s worth saying out loud what the umbrella is actually for, because that’s what decides where it goes next.

The point is the harness

The brief for every one of these libraries is the same: keep it small enough that a harness — a coding agent that may be partly or entirely a local model, see Le Harnais — can wire it into a working tool in one sitting. No framework to learn, no plugin system, a surface you can read end-to-end and audit. That constraint is the design, not a nicety.

It also explains the recent direction of the libraries that already exist:

• simple-network grew a post-quantum secure channel (hybrid ML-KEM + XChaCha20 + ML-DSA) and an optional Tor onion carrier — both behind feature flags so the base stays lean.
• simple-ui gained an AgentLog widget: a rendering target for streaming agent activity, because if a model is driving the tool you want to see what it’s doing.
• simple-secrets got a written threat model and fresh-secret TOTP minting.

Two new members, by the same logic

If the harness is going to run tools, two gaps follow directly.

simple-remote — remoting for the capability-scoped shell. Run a limited-shell session on another host over simple-network, under the same explicit resource and policy limits. No ambient SSH trust; just the scope you grant.

simple-sandbox — a boundary around the blast radius. When the harness lets a model run tools, you want side effects contained, filesystem and network access mediated, and an audit trail of what actually happened. Paired with limited-shell, what can run and what it can reach are both scoped.

Both are coming soon. The umbrella now has its own page laying out the libraries and the design rules — small over general, capabilities over trust, and public mirrors that carry just the code that compiles.