Performance
Built to stay light on every device
An agent that slows the device down defeats its purpose. Here is what the Hopix agents cost on our own devices, how we measured it, and the limits that keep that cost in check.
Measured
What each agent costs
The published agents, at their default level and with every measurement on, on our own devices.
Windows
- Memory
- 46–55 MiBprivate working set, the Memory column of Task Manager
- Disk
- 5.5 KB/s
- Device
- Windows 11 laptop, Intel Core i7 with 16 threads, 16 GB, on Wi-Fi
- Measured
- Oct 20–25, 2026 · 61 hours of use
macOS
- Memory
- 124–130 MiBfootprint, the Memory column of Activity Monitor
- Disk
- 3.3 KB/s
- Device
- MacBook Air with Apple M3, 16 GB
- Measured
- Oct 18–25, 2026 · 159 hours of use
Linux
- Memory
- 175–205 MiBresident memory (RSS), the RES column of top
- Disk
- 0.4 KB/s
- Device
- RHEL 9 virtual machine, 4 vCPU, 10 GB
- Measured
- Oct 25, 2026 · 15 hours
How we measured
Every agent reports, each minute, the CPU, memory and disk of the busiest processes on its device, itself included. CPU and disk come from that data, for one device per system on the dates shown. CPU is the average over every minute, as a share of the whole device; in the minutes when the agent was not among the busiest processes, we counted it at the level of the tenth one, so these averages are upper bounds. On macOS it includes the system log helper that the agent uses. Disk is the average of the minutes when the agent was listed. Memory is the one each system shows in its own process monitor: on Windows and macOS, from the median to the value of 95% of the minutes since the versions published on Oct 25, 2026 were installed; on Linux, where the agent does not report the resident memory that top shows, we measured it directly on the device that day, while idle. On macOS and Linux, memory grows during the first hours after the agent starts.
By design
What keeps the cost in check
A budget for the detailed measurements
On Windows and Linux, a CPU guard pauses the most detailed measurements when they pass 5% of one core over 10 seconds, and resumes them on its own. The rest of the agent keeps measuring meanwhile.
A ceiling per source
On Windows and Linux, each detailed source has a ceiling of events per second. Above it, the agent lets the excess go instead of doing more work.
Low priority on Linux
The Linux service runs with low CPU and disk priority and a memory ceiling of a quarter of the RAM, so the applications on the server come first.
Summaries, not traffic
The agent sends compressed per-minute summaries over one outbound connection. If the cloud cannot be reached, it keeps up to 50 MB on disk and sends them later.
Three monitoring levels
Full, events only, or nothing in the packet path, for each system or for a single agent. Lower levels measure less and cost less; our team sets the level you ask for.
Its cost, in your console
The console shows the CPU, memory and disk of each device with its busiest processes, the agent among them when it is, so you can check its cost on your own devices.
Under load
When the device is busy, the agent steps back
Heavy traffic makes the detailed measurements cost more. In our tests on the Windows laptop, a sustained download near 600 Mbit/s took them past the budget within seconds: the guard paused them for one to two minutes at a time, and they came back on their own once the load was gone.
See Hopix on your own network
Book a 30-minute walkthrough. We install an agent with you and look at real data together.