You cannot verify
what you cannot see.
SystemPulse is a real-time system monitor for Apple Silicon Macs. It surfaces the power, thermal, and performance truth the operating system does not expose: per-domain power draw, thermal headroom read from the kernel's own pressure level, per-process GPU utilization measured from the graphics driver, and a power workbench for local inference runs.
It also watches what coding agents actually do on the machine. That is the same discipline in a second register. Observability precedes verification, and before you can trust a machine that thinks, you have to see what it is doing, watt by watt and syscall by syscall.
macOS 14+ · Apple Silicon · No root required · No account, no telemetry
Download opens at release.
Every watt. Every degree. Every core.
SystemPulse reads directly from Apple Silicon hardware interfaces. Measured rails from SMC and IOReport and counters from the Mach kernel, presented at the precision the hardware reports them, with the few domains Apple does not expose on a rail labeled as modeled rather than quietly folded in.
Menu Bar
Power draw, temperature, and fan speed live in the menu bar at all times. Option-click cycles through four display modes. Ten additional widgets can be pinned individually, or merged into a single status item. Three density levels for machines where menu bar space is tight. Per-widget mini-dropdowns reveal detailed stats without opening the full panel.
Dropdown Panel
Eleven configurable sections, dragged to reorder or toggled off. An eight-domain power waterfall, with the domains Apple does not expose on a rail clearly labeled as modeled. A per-core CPU heatmap with cluster topology. Thermal rate-of-change by ordinary least squares, and a throttle banner driven by the kernel thermal-pressure level that warns one level before throttling begins. Pin it to the screen edge and it stays visible while you work.
Process Manager
Seven tabs. Power and temperature sparklines with top processes. A logical die and cluster view whose per-core temperature comes from a dedicated SMC sensor per core, not interpolation. A process table with ten sortable columns and a parent/child tree, per-process GPU utilization measured from the graphics driver per-client counters, and Security-framework code signing verification in the detail panel. An inference workbench that records DRAM, GPU, and CPU power against the tokens per second you report. Twenty-four hours of persistent history on disk, correlated against thermal throttle events.
Your agent has your keys.
A coding agent on your machine can read your SSH keys, your cloud credentials, your git history, and your environment variables. Right now. That is not a vulnerability. It is the design.
When a prompt injection hijacks an agent, it does not install malware. It uses the tools already authorized. Antivirus sees nothing. The firewall sees nothing. Only an external observer, watching what the agent actually does, can tell the difference between the work you asked for and the work someone else asked for.
Observe. Correlate. Respond.
Security monitoring is opt-in and off until you turn it on. When it is on it samples every ten seconds. We say so here because a monitor that overstates how continuously it watches is worse than one that does not watch at all.
Every process. Every socket. Every secret.
Destination IP and port for the TCP connections of your own processes. Secrets matching known key shapes detected in process command-line arguments. Changes to SSH keys, cloud credentials, keychains, and package registry credentials observed as they happen through FSEvents. Coding agents identified automatically, including Claude Code, Cursor, Copilot, Windsurf, and Aider.
No single rule catches a multi-stage attack.
Anomaly rules score each process independently, and deterministic correlation patterns surface what individual rules cannot. A sensitive-file alert and a network alert from the same detected agent within thirty seconds flags a possible exfiltration pattern. A secret-exposure alert and a network alert within sixty seconds flags a possible credential-theft pattern. The logic is pure and offline, so there is no model to talk out of its conclusion.
Block it before it runs, not after.
SystemPulse listens on an owner-only Unix socket that the Claude Code PreToolUse hook connects to. With that wrapper installed, agent shell commands matching pipe-to-shell, eval, or a recognized credential shape are refused before the command runs rather than reported afterward. System notifications for security events, signature-anchored trust overrides that stay reviewable and revocable in a ledger, and a security timeline exported as JSONL for whatever reads it next.
Engineered properties, not promises.
External Observation
A compromised agent cannot monitor itself. SystemPulse runs outside the agent process tree and reads the same kernel data structures Activity Monitor reads, so the agent is not in a position to switch the observation off from inside its own process. Both systems have to fail independently for an attack to land.
The observer is not a participant.
Behavioral Detection
Signatures detect threats that are already known. Behavior detects the ones that are not. SystemPulse scores what a process does rather than what it claims to be, so there is no signature database to maintain and no definitions to download.
What it does, not what it is.
Direct Hardware Reads
No Electron wrapper. No Python underneath. Every sensor reading comes from a direct hardware or kernel API rather than from scraping a command line tool: SMC, IOReport, the Mach kernel, the AGX graphics driver, and Darwin notification APIs, written for Apple Silicon rather than ported to it.
Built for the hardware it measures.
If you trust your agent completely, you do not need this.
If you do not, you need something that watches from outside the thing being watched.
SystemPulse is distributed directly by PhysOpsAI. It is not in the Mac App Store, because the sandbox the store requires forbids the hardware reads the app is built on.
There is no account and no telemetry. The app collects nothing about you and sends nothing about you anywhere. Two things do touch the network and both can be turned off: it checks a cryptographically signed update feed, and it resolves hostnames for the connections it shows you.
No App Store. No account. No tracking.
Download opens at release.