Fumble

Fumble is a free, open-source typing coach for macOS that learns from ordinary typing across the day. It lives in the menu bar, observes physical key events and timing in apps such as editors, terminals, and email, then identifies the keys and key pairs that are costing the user the most time. That makes it different from a typing website that can only judge a short exercise performed inside its own page. Fumble starts with the habits a person already has and turns those measurements into focused practice. It is especially relevant to developers and other heavy keyboard users whose everyday text includes commands, symbols, identifiers, and patterns that generic prose drills rarely cover.

The ranking is based on time lost rather than raw slowness. Fumble compares key latency with a baseline derived from the user's own timing, then accounts for how often each key is pressed. Its baseline uses the median of per-key high-percentile latencies rather than pooling every keystroke, so a common slow key cannot raise the benchmark and hide itself as easily. A mildly slow key used thousands of times can therefore rank above a very slow key that appears only occasionally. The same analysis covers transitions between adjacent keys, where a difficult pair may be costly even if each key seems fine in isolation. Same-finger pairs are shown but kept out of drills because their physical constraint isn't necessarily a habit that more repetition can fix. Correction rate is tracked separately so speed and accuracy problems don't get collapsed into one score.

Its words-per-minute figure is designed to be inspectable. The app excludes software-injected events from text expanders, automation tools, remote control software, and other processes. It also leaves out key autorepeat, idle time, long thinking pauses, and activity inside Fumble's own practice window. The pause threshold adapts to the user's timing for each transition instead of applying one flat cutoff to every key pair. Pasting and most editor completions do not create ordinary key events, so they do not inflate the count either. A dropdown reports what was excluded rather than hiding the filtering behind a single unexplained number. This helps a user understand whether the daily metric reflects finger movement or an unrelated source of input.

Practice is more than one weak-key exercise. The adaptive Trainer begins with information captured from real typing, so keys already typed confidently can start further along while an actual weakness receives attention from the first session. An on-screen keyboard visualizes confidence, and per-letter views show recent performance, best performance, and lesson speed. Weak Spots generates continuous real-word drills weighted toward slow keys and transitions. Custom Text accepts pasted material, Numbers serves digit groups, and Code uses symbol-heavy pseudo-code fragments. Users can choose stop-until-correct or advance-through assistance, show whitespace dots, select cursor behavior, enable key and error sounds, and follow a daily goal with streaks. Practice results do not feed back into daily capture and distort the model.

Because system-wide key observation is sensitive, Fumble makes narrow data collection a core design decision. It records physical keycodes and timings, not the characters those keys produced, and writes counters and latency histograms rather than a replayable keystroke sequence. It stops recording when a password field has focus. The project discloses that adjacent-pair frequencies remain a limited residual risk and documents how users can inspect the stored data. There is no account, telemetry, or network code. The source is public under the MIT license, and a command-line tool can print today's summary, all recorded days, or the raw stored data for inspection.

Fumble requires macOS 14 or later and the Input Monitoring permission. That permission is necessary for all-day analysis and also prevents distribution through the sandboxed Mac App Store. The direct release is Developer ID signed and notarized, while installation is also available through a Homebrew cask or from source. The project is explicitly early, with a small community and some real-world combinations still to be tested. It won't suit Windows or Linux users, and people uncomfortable granting system-wide input access may prefer a browser typing tutor. For Mac users who accept that permission, it offers an unusually transparent link between actual typing friction and the exact drills used to address it.