What the app can do today – and what it cannot yet.

Available · beta

Selection and actions

AYVI captures selected text through public Accessibility APIs, as far as the app exposes it. Local heuristics detect code, JSON, SQL, error output and several programming languages, among others. Quick Mode shows up to four recommended actions; Deep Mode offers the full action bar and a free question.

Recommendations are based on locally detected content, language rules, pinned actions and usage counters. No cloud classification is used. Only action identifiers and counters are stored, never selected text.

Read-only text: When a selection cannot be changed, for example text on a web page, AYVI offers only analysing actions (explain, summarize, translate, and for code also debug and tests). Actions that change the text, and replacing, are not offered there. How reliably AYVI recognizes “cannot be changed” depends on the role the app reports through Accessibility; this has not been checked app by app.

Writing in German and English

For detected prose, minimal spelling/grammar correction and rewriting come first. Correction is meant to keep language, wording, meaning, tone, names, numbers, quotes and formatting. Before rewriting you can choose style, length, output language and an optional instruction; confirmed settings stay local as a preset. The optional explanation of changes is separate and never mixed into the replacement. Results are reviewed as a word diff and applied only after confirmation.

Providers

Included are Apple Foundation Models on supported macOS 26+ systems, a labelled Mock/demo provider, Ollama, Anthropic and OpenAI-compatible endpoints. When the system model is available, Apple is preselected on fresh installs; otherwise Mock remains the fallback. The local Foundation Models streaming path was smoke-tested on the development Mac with synthetic input. That checks neither answer quality nor the UI nor other devices. Cloud providers and their live compatibility have not been tested.

Diff, replace and undo

Normal rewrite actions show a visible diff and change the selection only after explicit confirmation. Before changing anything, AYVI checks process, app, focused element, selected text/fingerprint and – where available – the range again. Failures keep the result copyable and can offer a fresh capture.

Correct and Continue: Once the answer is in and the selection can be changed, AYVI opens the diff review by itself; nothing is written until you confirm. “Continue” inserts the continuation after the selection and never replaces it, because the answer contains only the continuation. For read-only text no review opens.

Direct Improve is the only exception: after this deliberately chosen, clearly labelled action the app may replace the selection without a second diff confirmation. For prose it only requests minimal correction and never starts a style rewrite. The setting is on by default because the user explicitly asked for it. With “Direct Improve without confirmation” switched off, the action goes through the normal diff and confirmation. Its own shortcut is off by default. Undo is offered; the clipboard path has additional limits.

Further available features

• Local conversations, on by default, with search, resume, rename and delete. A first-launch notice explains storage; selected text and conversation storage can be switched off separately. The archive is encrypted (ChaCha20-Poly1305, key in the keychain).
• Local secret scanner before every cloud request and before saving; “Mask all” with placeholders such as [API_KEY_1] is the default. Heuristic: it can miss secrets.
• Optional detailed privacy preview before cloud requests.
• Experimental, off-by-default screen region capture with local Vision text recognition, an editable confirmation preview and a read-only OCR selection.
• Experimental, off-by-default code lens markers from a structured explanation and current public Accessibility line bounds; the same explanation is available in the panel as a keyboard- and VoiceOver-friendly list.
• Experimental Services menu entry “Explain with AYVI” for apps without Accessibility support: read-only, no clipboard.
• Experimental support level “Coach”: guiding questions instead of a solution, never offered as a replacement.
• Experimental app profiles, off by default: matching actions first, preselected support level, cloud blocked per app. Profiles never choose a provider.
• Experimental, off-by-default compile check (javac, swiftc, py_compile) that never runs the code.
• Experimental, off-by-default voice input while holding the shortcut. Local recognition is enforced; the editable transcript is never sent automatically.
• API credentials in the macOS keychain; cloud providers with a visible disclosure; “Only local models” blocks cloud providers.
• Built-in self-test window and a per-app compatibility check that shows capabilities only, never content.
• Selection preview, free questions and local Java shortcuts in the AYVI question field.

Java learning path · experimental, off by default

When enabled, AYVI detects Java concepts with local, deterministic heuristics limited to 65,536 characters. Comments and literals are blanked first; the detection is still not a parser and not reliable proof of knowledge. Only explicit actions change the status. Explanations and exercises are stored locally, reviews follow growing intervals. Only concept IDs, status, timestamps and counters are stored and exported. Prompt context is a separate opt-in; when a cloud request contains such concept names, AYVI shows them first.

Important limits

• App compatibility of Accessibility capture/replacement, shortcut delivery, keyboard and VoiceOver have not been systematically checked by hand – see compatibility. ScreenCaptureKit permission, real screen capture and Vision recognition were not tested in practice either.
• The secret scanner is heuristic: it can miss keys, tokens or passwords or flag harmless text, and it does not reliably detect personal data in general.
• Models and target apps have finite context and selection limits. Chunked processing handles sections independently and guarantees no context across section boundaries; the result must be reviewed.
• There is no cursor/line/function context without a text selection and no reading of the whole file.
• There is no local code execution. The optional compile check only checks syntax and types and runs nothing; compiles does not mean correct.
• Lens markers have not been checked in real editors (position accuracy, scrolling, VoiceOver). The markers themselves take no focus; keyboard and VoiceOver users use the list in the panel.
• A local development bundle is ad-hoc or locally signed and not notarized; there is no public binary release. A release script for Developer ID signing and notarization is prepared but has not been run.

Planned or under review

A safe local code runner stays blocked until a supported isolation provably prevents network, home-folder and persistent file access. Contextual cursor selection, PEEK mode, verify, hunk selection and custom actions are not current features. Details are in the roadmap. Planned features are not simulated in the interface.