How OpenCues compares with the tools closest to it: AI keyboards, dictation apps, system-wide writing assistants, browser assistants and developer tools. Reviewed July 2026. Traction figures are public reports or vendor claims.
1. OpenCues voice-mode is text-to-speech output (it reads tips aloud), not voice input. Dictation tools pair well with OpenCues: they put text in the buffer, OpenCues works on whatever is there.
2. Agent tasks (agentically X _) operate on your text buffer rather than commanding external apps.
3. Apple and Windows ship OS dictation separately from their writing tools; capabilities are counted per product.
4. Terminal integrations are native apps; this column tracks arbitrary GUI apps such as mail clients and editors. OpenCues reaches these system-wide on Windows through UI Automation (Notepad, WordPad, the browser address bar, Explorer, Win32 / WinForms / WPF fields), integrated into each field's text pipeline rather than coarse OS injection (see 6), so cues, cycling and revertable edits keep working.
5. Fully local tools avoid sending PII by never calling a cloud provider; this column tracks redaction on cloud calls.
6. "Any app" reach comes from OS-level injection: clipboard swap and paste, synthetic keystrokes, accessibility APIs, OS text views, or screen OCR. Those routes offer coarse insert and replace verbs with no durable model of the buffer, which is why the per-word columns (cues, revertable edits) stay empty on the same rows. OpenCues integrates inside each host's text pipeline instead: richer state, narrower coverage.
Placements are qualitative, derived from the verified capability grid above. Hover any point for its name.
How much of your computer a tool covers, against how deeply it holds your text. OS injection buys breadth with coarse verbs; in-pipeline integration buys per-word state.
Whether you choose the model (keys, your existing plan, local), against how close the answer lands to where you were typing. espanso and Click to Do sit out: no model in one, no text edit in the other.
Horizontal position is measured: the average of the two explicit-task step counts from the table below, fewer steps further right. Vertical is the passive-layer tier: nothing volunteers, ghost text proposes next words, underlines flag problems, cues offer revertable alternatives to what you already wrote. espanso sits out: no model behind it.
OpenCues is the only row that is simultaneously multi-surface, overlay-free typed trigger, model-agnostic including subscription reuse and local inference, and permissively open source.
How many user actions each tool's documented happy path takes. A step is one discrete action: a selection, hotkey, click, typed prompt (one step regardless of length), app switch, paste, or confirm. Automatic insertion counts zero. Counted from each product's official documentation, July 2026. The third column is the asymmetry the first two hide: a passive layer surfaces problems and alternatives on its own, at zero asking cost. Without one, you either run the rewrite flow on suspicion or ship the weakness without knowing. The next table splits the same journeys at the moment the result first exists. Step counts measure the cost of one interaction, not its range: two tools can pay the same steps for very different capability, which is what the grid above measures.
1. espanso expands a snippet you saved earlier, not an ad-hoc AI answer.
2. BoltAI documents the trigger reading the field's content; rewrite-in-place of existing text is not explicitly documented.
3. Wispr Flow's Command Mode is a paid-plan feature enabled once in settings.
4. Copilot's two-step path is code completion inside an IDE.
5. Monica's toolbar does not document reinsertion; counted as copy then paste.
6. Ghost text proposes the next words before they land (accept-first); cues propose revertable alternatives to words already written. Both count as applying one automatic suggestion.
The same fourteen journeys from the table above, split at the moment the result first exists. The left column is the ask; the right column is the end-lag: every gesture between the result existing somewhere and you being back at your caret with it in place. The two columns sum to the totals above.
Both axes measured from the split above: steps to ask, and steps back to work. Clustered dots share exact coordinates. espanso is plotted from its snippet flow; there is no model behind it.
What it costs to reject an AI change you did not want, before it lands and after. The after column is the one that matters: once a paste is plain text, undoing one part of it without losing the rest is manual work everywhere except where edits stay individually owned.
Seamless means no side-effects on the workflow you already have: nothing new to look at, dismiss, or learn per platform. A count of the surfaces each tool introduces during the flows above.