Guardian at github.com/jdetle/guardian is a macOS stack for AI coding agents (Cursor and OpenAI Codex today): sampling, hooks, and policy so the host’s real load shows up in every session.

Status — AI draft: I have not done my own pass on tone, claims, or emphasis yet. Treat this as generated scaffolding; I’ll revise under Human when I’ve actually read it with coffee.

What it actually is

Guardian is an agent resource monitor: a Rust daemon (guardiand) samples CPU, memory, swap, thermal state, Docker load, Cursor process RSS, and disk pressure on your home volume, then writes atomic state under ~/.guardian/ (notably state.json and hook_policy.json). Editor hooks inject that reality into every session so the model isn’t flying blind when the host is already on fire. Optional prompt gates can block sending a prompt while pressure or Cursor memory is past policy — with explicit human-in-the-loop escape hatches (proceed_once, snooze, zeno), so “blocked” never means “bricked.”

That’s the product in one breath: observable pressure + honest context for agents + daemon-side enforcement + optional gates with resume, pitched at people running agents on real laptops (tight RAM, full disks, thermally sad CPUs), not infinite-cloud fantasy hardware.

Why that’s worth something

Cursor on an older MacBook or 8 GB of RAM is not the same product as Cursor on a workstation with headroom. Agents still want parallel shells, subagents, Docker, indexers, and long chats. When the machine is already swapping, completions time out and the UI stutters — not because the model got dumber, but because the host ran out of air. Guardian makes “how hard are we pushing this machine?” legible to both you and the tool chain, instead of invisible until the fan sounds like a jet.

For a company, the pitch isn’t “install a cool daemon.” It’s: if you’re mandating AI-assisted development on a standard-issue fleet, you’re implicitly mandating load you don’t measure. Help-desk tickets, “my IDE froze,” and people learning to work around the official stack are all symptoms of the same gap. A shared pattern — pressure levels in hook output, optional gates so nobody floods prompts into a box that’s already critical, disk and fork guards, optional Docker throttling — is how you keep agents useful without turning every laptop into a stress test. That’s operational hygiene for agent rollout, not a nice-to-have.

What ships in the repo (high level)

The README lays it out in a table: guardiand for sampling and policy; a guardian CLI for snooze / once / zeno; Cursor hooks (session start, before submit prompt, before read file, pre-tool use, subagent start, …); optional Codex hooks; config.toml for thresholds and gates; a deferred work queue (JSONL + CLI) when prompts can’t run yet; an optional SwiftUI menu bar app; stress tests in Docker so validation doesn’t require melting your own machine.

None of that replaces your security team or your cloud posture — it’s about local honesty: what the machine is doing, what Cursor is holding in RAM, whether disk is about to make everything worse, and whether the next prompt is a good idea.

Open source names collide — “Guardian” appears in several repos. The value prop here is keeping agents and modest Macs from working against each other, with hooks and policy you can audit. Evaluation starts from github.com/jdetle/guardian and the install path in the README.

AI draft: This entire version is machine-generated. I will replace or rewrite it after human review — facts, emphasis, and whether “big corps should mandate X” is even the right frame. Do not treat it as final author voice or a promise about anyone’s IT policy.