A native Windows task manager with the classic Windows 10 layout, live hardware graphs, process controls, search, and light and dark themes. Written in C++20. Runs as a single executable without .NET or additional runtime downloads.
See the v1.3.0 release notes for scoped search, CSV export, sampling recovery, reliability fixes and the Python development tools.
Requires 64-bit Windows 10 or Windows 11 and administrator approval. ARM64 and 32-bit builds are not provided. Windows 11-specific features have not all been verified on Windows 11.
- Download
TaskManager.exefrom the latest release. - Put it somewhere you want to keep it, then run it. There is no installer for standalone use.
- Approve the Windows administrator prompt. The app does not launch if you cancel.
- On first launch, choose No at the replacement prompt if you only want to try it alongside Windows Task Manager.
The executable is currently unsigned. Windows may show a security warning. Check that the download came from this repository; each release includes a SHA-256 checksum file. Run certutil -hashfile TaskManager.exe SHA256 in Command Prompt and compare it with SHA256SUMS.txt.
Settings and locally recorded CPU history are stored in %LOCALAPPDATA%\TaskManagerNative.
Enable Options > Replace default Task Manager to redirect Windows Task Manager launches, including Ctrl+Shift+Esc, to this app. This affects all accounts on the computer. Setup installs a protected copy at %ProgramFiles%\TaskManagerNative\TaskManager.exe; moving your downloaded copy does not break the replacement.
Turn the same option off to restore Windows Task Manager. Setup does not overwrite another application's replacement, modify Microsoft's executable, or install a background service.
If you need to restore Windows Task Manager outside the UI, run this from an administrator Command Prompt:
"%ProgramFiles%\TaskManagerNative\TaskManager.exe" --disable-replacementUpdate: for standalone use, close the app and replace your downloaded executable. If replacement is enabled, disable it first, close the installed app, delete its old executable, then run the new download and enable replacement again.
Uninstall: disable replacement first if enabled, close the app, and delete the executable and its installed directory. Delete %LOCALAPPDATA%\TaskManagerNative too if you want to remove settings and recorded history.
- Processes: grouped apps, background and Windows processes; expandable process groups; CPU, memory, I/O, network and GPU activity; end tasks and restart Explorer.
- Performance: CPU and per-logical-processor graphs, memory usage, disk activity, network throughput and GPU engines. Available hardware details depend on Windows and the device's drivers.
- App history: locally recorded CPU time, with a reset option.
- Startup: Run entries, startup folders, boot/logon scheduled tasks, automatic services and drivers, and packaged startup declarations. Supported Run and folder entries can be enabled or disabled; other sources link to Windows management tools.
- Users, Details and Services: session activity, process metadata and controls, service status and start/stop/restart actions. Failed service refreshes retain the last known rows and show an unavailable or outdated status with the Windows error code; F5 retries immediately. Windows protection and permissions still apply.
- Appearance: Light, Dark and Use Windows setting; High Contrast follows system colours. Columns and sorting are saved separately for each tab.
- Search: filter table rows by text or combine exact PID, name, publisher, command line, path, account and status fields. Quoted values and field exclusions are supported. Matching groups retain their context. Search does not request a fresh system sample.
- Export: save the currently filtered and sorted table to a UTF-8 CSV file with File > Export current table or Ctrl+Shift+S. The export includes the displayed columns in their current order, with complete cell text.
Right-click a column heading to choose visible columns. Ctrl+F opens search; Enter applies it immediately and focuses the results; Esc clears it. F5 refreshes. Ctrl+Tab cycles tabs, and Alt+1 through Alt+7 selects one directly. Double-click a Performance graph or its sidebar to enter a summary view.
View → Update speed → Adaptive (low overhead) targets process updates between 250 ms and 4 seconds, backing off when measured sampling and table-update work becomes expensive. It aims for roughly 2% sampling/display wall-time duty, not a guaranteed CPU cap; asynchronous metadata, painting and tracing also use resources. Adaptive uses a 1-second target on other full-size tabs. It is the default for new settings; existing saved update speeds are preserved. Fixed High targets 500 ms, Normal 1 second, and Low 4 seconds.
Process CPU, memory, I/O, window classification and traced network activity are sampled on each process update. More expensive detailed memory, per-core, disk, network-interface and GPU counters normally refresh no more than once a second during automatic polling, then are reused between process updates. Processor-topology changes can force an earlier refresh. Their rate calculations use their own sampling interval, and cached process GPU data is matched by PID and creation time. F5 also invalidates the resource cache.
Sampling time is included in the cadence rather than added to it; slow samples cannot trigger a catch-up polling loop. Minimized monitoring slows to at least 4 seconds, and restoring the window consumes the latest snapshot and requests a fresh one unless paused. Holding Ctrl temporarily freezes table updates; releasing it consumes the latest available snapshot. These are sampled measurements, not instantaneous guarantees: CPU and I/O rates describe the interval between snapshots, and short-lived processes can exist entirely between samples. Adaptive backoff is tested with simulated costs, but performance on physical 20-year-old hardware has not been verified; the Windows 10/11 x64 requirement still applies.
If a live sampling pass fails, the app keeps the last complete snapshot and changes its title to Task Manager - Live updates unavailable; retrying.... It retries after 250 ms, doubling the delay after repeated failures up to 4 seconds. F5 requests an immediate retry, and closing the app interrupts the wait. A successful sample restores the normal title and the selected update speed; if updates were paused, they return to paused after recovery. The retained values may be outdated while the retry status is visible.
Stable tables, including filtered results, repaint only visible rows whose text, heat colours, icons or tree appearance changed. Numeric sorting keeps the viewport position instead of scrolling after a process whose rank changes. Selection and independent keyboard focus retain their process identities as rows move. Light and dark modes share this refresh path; changing appearance does not request another system sample.
Plain text remains a case-insensitive substring, including spaces: Orbit Browser searches for that phrase. Windows paths and command switches also work as ordinary text. A recognized field prefix enables queries with multiple terms; every term must match. Quote values containing spaces and prefix a field with - to exclude it.
| Query | Matches |
|---|---|
pid:4242 |
Exactly PID 4242. Plain 4242 still searches substrings. |
name:browser publisher:"Orbit Labs" |
A process whose executable name or description contains browser and whose publisher contains Orbit Labs. |
path:"C:\Program Files\Orbit\" -status:suspended |
Processes with that path text that are not suspended. Backslashes remain literal inside quotes. |
command:"--profile Main" |
Available command-line metadata containing that phrase. |
user:alice status:"not responding" |
A process associated with an account containing alice that is not responding. |
Fields are name, pid, publisher, command, path, user and status. They use the current snapshot and available metadata; an unavailable field cannot satisfy a positive term. Startup, service and history entries use the corresponding fields where available. User rows can also match the session account and status. For process groups, one member must satisfy the whole query. The group and its visible children remain together, so some context rows may not match individually. Group headings retain their unfiltered totals. Performance resources are not filtered.
Export captures the rows and displayed values when you invoke it, including expanded group rows. Category headings and hidden columns are omitted. Refreshes while the Save dialog is open do not change that snapshot. Saving runs in the background; the existing destination is replaced only after the new file has been completely written.
CSV files use UTF-8 with an encoding marker, quoted fields, and CRLF record endings. Quotes and embedded newlines are preserved. Formula-like cell text is prefixed with an apostrophe for spreadsheet import; embedded NUL characters are replaced with the Unicode replacement character. These are display-value exports rather than a raw numeric data feed. Export is available on table tabs with results in the full view.
This is an independent implementation, not a complete copy of Microsoft's Task Manager.
| Area | What to expect |
|---|---|
| Disk activity per process | Read/write transfer rate includes some non-disk I/O; it is not Windows' disk-only trace. |
| Power usage | An estimate from activity, not Windows' energy estimator. |
| App history | CPU time recorded while this app runs. No Windows SRUM network or tile-update history. |
| Startup | Broader inventory than the ordinary Startup apps list, but not exhaustive Autoruns coverage. Startup impact is Not measured. Some sources are read-only. |
| Memory and GPU details | Memory composition uses in-use/cached/free categories. Some stock GPU fields, including DirectX feature level and hardware-reserved memory, are absent. |
| Packaged apps | Grouped by package family, but classified as apps only when the group has a visible app window. Names and icons use available executable metadata, with a generic icon fallback rather than resolved manifest display names and logos. |
| Process controls | No UAC virtualization toggle or kernel dumps. Affinity covers one processor group, up to 64 logical processors. Protected processes may refuse actions even with elevation. |
| Efficiency mode | Available only when Windows supports querying the process's policy. Requests EcoQoS and lowers priority; it is not a CPU cap. Disabled on the tested Windows 10 build 19045. Live Windows 11 verification is still outstanding. |
| Search and themes | Added to the classic tab layout. Search supports plain substrings and scoped AND/exclusion queries, without OR, regular expressions or Performance-resource filtering. Popup menus and system dialogs remain OS-dependent. |
See feature differences and startup coverage for details.
On Windows x64, install Visual Studio 2022 or newer (or its Build Tools) with Desktop development with C++, a Windows SDK, CMake 3.24 or newer with support for your Visual Studio version, and Python 3.10 or newer. Run this from the repository directory in your preferred terminal:
python tools/taskmgr.py build
The command builds the production executable, runs self-test, sampler and UI/stress CTest cases in a separate developer build, checks the production manifest and diagnostic exclusion, and copies the verified binary to artifacts/app/TaskManager.exe. The tooling uses Python's standard library; there are no Python package dependencies. Run python tools/taskmgr.py --help for benchmarks, paired performance comparisons, compiler studies and packaging commands. Python is needed only for development, not to run the app.
Production requires elevation and excludes test, benchmark and screenshot tooling. Developer builds run without an elevation requirement and must not be distributed as releases. When the build command runs unelevated, it verifies Windows' refusal to launch production without elevation; runtime command-rejection checks require an elevated test shell.
Bug fixes, usability improvements and performance work are welcome. Open an issue with your Windows build, reproduction steps and expected behaviour, or send a focused pull request.
For UI changes, include a screenshot and check light, dark and High Contrast modes. For performance changes, include before/after measurements from the same machine. Avoid posting process command lines, user names or startup reports without reviewing them for private information.
Read CONTRIBUTING.md for build, test and benchmark commands. A project license has not been selected yet.