Select text or a screenshot
A screenshot you select is converted to editable text with bundled local Latin OCR. This is a deliberate import, rather than automatic reading of incoming messages in Teams, Gmail or other apps.
Consierge in detail
An assistant that adapts to you.
Dictate, express, reply. With local models, your own style and clear choices about which context is used.
01 / Dictation
Speak without losing your train of thought.
Tap to start dictation; holding lets you record and finish in one gesture. The waveform makes recording visible. The X discards it, including by dragging and releasing over the cancel area.
Parakeet recognises speech locally. Recordings are limited to five minutes and recognised in 15–20-second segments. Optional German cleanup handles up to 24,000 characters in sentence windows. Oversized sentences or capped suggestions retain the original text. Segment previews are not immediate token streaming; you review the result.
- Dictation in enabled, suitable app text fields
- Outlook subject and message tested on the S24
- Adjustable button size, opacity, position and pause
Text workshop
Deliberately revise text. Choose the result yourself.
In the internal Android beta, work on deliberately imported text: clean up, rewrite, translate, compose a prompt or prepare a reply. Choose language, instruction, route, paired server and model per mode.
Review changes
Optional dictation review shows selectable changes before insertion. A recent insertion can be undone briefly only while the editor, session, text and cursor still match the recorded state.
Your approval remains essential
Review every suggestion. The text workshop and responder never send messages automatically. A deliberately selected import extends possible uses, but does not provide a native message or thread adapter.
02 / Writing style
Your own terms. Useful formatting rules.
Assign six styles to apps or categories. They mainly change formatting details such as the final full stop, selected lowercase words or hashtags. They do not turn casual dictation into a completely rewritten business email. Additional apps require explicit permission; only suitable text fields are supported.
Optional style learning considers evidenced formatting habits. Up to twelve approved examples of your own writing can also influence wording and sentence structure in replies and the text workshop. They provide model context without training model weights. Automatic collection of your wording is a separate option, off by default.
Your dictionary and snippets
Save names, brand terms and your own spellings. Text snippets expand a spoken shortcut into saved text. Search helps you find and edit entries.
03 / Replies
A suggestion. The last word is yours.
With your permission, Consierge uses a deliberately selected readable message or a window of 1, 3 or 5 visible messages in a supported messenger. A selected quote takes priority. Unclear or truncated context may require explicitly sharing the text. Your short instruction guides the draft; you review it before insertion.
Neutral and Warm are enabled initially; you explicitly enable other tones. Per profile, choose reply length, intensity, boundaries and one to three drafts generated in sequence. Free-form local replies and rewrites still take seconds, depending on the model and device.
Consierge prepares drafts. It does not send messages on its own or universally identify people by name.
04 / Memory
Keep context. Keep control.
Conversation memory is separate from style learning and is off by default. It stores limited, enabled excerpts encrypted. You choose a retention period of 1, 7 or 30 days.
Search deliberately imported conversation excerpts, review summaries and return to earlier revisions. You confirm core notes yourself. Password-protected transfer and selective deletion are available; imported profiles require renewed consent. Consierge does not automatically scroll through messenger histories.
What does the responder do?
The experimental responder requires global and profile-level consent. In WhatsApp, automatic matching is limited to a verifiable adapter key. For Signal, Telegram or an ambiguous identity, you explicitly select the current input session. Display names do not identify a person. It only prepares drafts; insertion and sending remain your actions. Real messenger navigation still needs acceptance testing.
Open paths forward
From your device to your own AI infrastructure.
On-device processing is the default route. An optional paired Ollama gateway on your LAN is implemented in the internal Android beta. LiteLLM is a separate, deferred path.
On your device
Parakeet for speech, rules and local models for text. The tested scope refers to the S24 Ultra.
Ollama
Pair your own Ollama gateway and choose cleanup and reply models separately. Select phone only, prefer LAN with one on-device fallback, or LAN only with visible failure reporting. Speech recognition stays on the phone.
Official documentation ↗LiteLLM
A separate experimental path for your own model providers and gateways. Production validation and release remain open; this is not a finished cloud service.
Official documentation ↗Android 0.4.0-beta.1 was internally qualified on the S24 Ultra: DE/EN UI, local OCR, data controls, text modes, segmented speech recognition and paired Ollama LAN. Tests used synthetic own data. Real messenger workflows, microphone interruptions and other devices still need separate testing.
Transparent development status
Real screens. Honest limits.
The views shown come from an earlier native Android prototype on a Galaxy S24 Ultra. The memory profile is synthetic. Initial native previews for other platforms do not yet match Android features. Model responses can be wrong; review them before use.
View Android status →