Features
More ways to make it yours.
Chat, documents, images and voice. Everything you can do with Lapis, organized around the way you work.
Take a conversation further.
From a first draft to a different direction, without losing the work before it.
Projects and instructions
Group chats by project, give each its own instructions and adjust response tone. Projects require PRO.
Duplicate and fork
Copy an entire conversation or branch from a message. The original stays intact; creating a branch does not automatically generate an answer.
History, search and drafts
Return to saved chats, search your history and keep your draft when moving between screens.
Your work, in other formats
Export as PDF, Markdown or text. Copy formatted code and regenerate an answer with another compatible model.
Revise the message. Continue from there.
Edit a sent message and confirm Save & Regenerate to replace later turns. Attachments are preserved; duplication and branching remain separate actions.
Bring your documents into the conversation.
Text, tables or a scanned page. Start with the material you already have.
PDFs and text files
Attach PDF, TXT, Markdown, CSV, JSON, XML and YAML with PRO. The content must fit the selected model’s context window.
Local text recognition
Lapis reads text in scanned PDFs with Apple Vision, including for text-only models. Recognition can make mistakes; it does not interpret diagrams or page layout.
Questions about photos
Add an image from the camera, Photos or Files. Visual understanding requires PRO and a compatible vision model.
Clear import limits
Up to 32 MB per document and 2 MB of extracted text; scanned PDFs are limited to 40 pages. Oversized content is reported rather than silently truncated.
From an answer to the page.
Open PDF page references in the original document, including OCR scans. The model proposes the citation: check that the page supports the answer. This is not automatic fact-checking.
Drop in what you’re working on.
Drag documents or photos into chat on iPad and Mac. They are prepared in the draft, never sent automatically, with the same attachment limits and requirements.
From an idea to an image.
Create, refine and keep images in the same workspace.
Bonsai Image 4B, locally
Generate with PrismML Bonsai in square, portrait or landscape. Binary and Ternary variants offer a size/quality choice; model downloads and sufficient memory are required.
Apple Image Playground
On supported devices, choose Apple’s native interface from the image selector. Its controls and availability are separate; offline operation is not guaranteed.
A gallery that stays with you
Keep creations locally, zoom and crop, copy the original prompt, regenerate, share and select multiple images for batch actions.
Choose a model. Know its limits.
Available memory matters more than a list of model names.
A device-aware catalog
Liquid, Mistral, Qwen, Llama, Gemma and other supported families. See capabilities, size and memory guidance; import compatible Hugging Face models with PRO, including uncensored variants.
Apple, on-device or in the cloud
Apple’s on-device model depends on system readiness. Apple Intelligence Cloud is a separate online choice requiring consent, internet, compatible software and Apple eligibility, with its usage limits.
Verified downloads
Resume downloads, follow progress and opt into completion notifications. File verification helps prevent incomplete models from appearing ready.
Memory-aware context
Space is reserved for the reply and recent turns take priority when context fills. Saved history stays intact: this is not automatic semantic summarization.
Inference controls
With PRO, adjust sampling, response limits, seed, stop strings and cache on compatible models. Prefix reuse is bounded and isolated; KV compression is model-dependent.
Beyond the keyboard.
Voice and native integrations that fit the way you work.
English Live Voice
Talk with the downloaded Liquid model, local recognition and spoken responses. Voice supports English; PRO, compatible hardware and microphone/speech permissions are required.
Composer dictation
Dictate without opening the keyboard and review the text before sending. Read-aloud can be stopped from the message itself.
Widgets, Siri and Shortcuts
On iPhone and iPad, open chat or Live from widgets and use actions to ask, summarize, translate and rewrite. Background text utilities use Apple’s on-device model when available.
A native experience
Share text from other apps, use keyboard controls and adaptive iPad/Mac layouts, light and dark appearances and accessible text sizes. The app interface is in English.
Heart. A new local voice.
English speech uses Paradee v1.1 with the Heart voice. Live offers Liquid, on-device Apple Intelligence or Apple Intelligence Cloud where supported, without silently switching providers. Cloud needs internet and consent.
Fewer steps from idea to action.
On iPhone and iPad: paragraph/bullet summaries, rewriting styles, checklists and downloaded-model selection in Shortcuts. On Mac: open chat, voice, images or models through an App Intent. Subtle feedback marks dictation start and stop.
Everything in its place.
Task-focused settings, advanced options on demand and an About section with version, updates and useful links. An optional Mac menu-bar item lets you check and control the service without opening the window.
Privacy you can understand.
Where a request is processed should never be a surprise.
Local by choice
Downloaded models process on your device; chats and images are stored locally. Local AI needs no account or external API key.
Temporary and protected chats
Use temporary conversations or protect saved chats with Face ID, Touch ID or system authentication, depending on your device.
Your usage, without your messages
View historical tokens, weekly/monthly periods and activity. API cost comparisons are estimates, not bills or guaranteed savings. The ledger stores local counts, not chat contents.
Performance at a glance
Check response speed, memory and device status. Detailed performance and advanced controls stay out of the conversation’s way.
A Mac. Also a server.
The Mac implementation shares models between the app, Link and CLI. No Python installation or open window is required.
Your Mac, from the sofa.
Pair your iPhone or iPad by QR or code with a Mac on your local network. The Mac runs an installed text model and streams the answer back. One saved Mac and one served model; the Mac must stay awake.
Lapis, in your terminal.
Optionally install the lapis command from the Mac app. List and install supported models, send prompts and control the server or Link. It shares the app’s models and configuration, without another engine or automatic shell changes.
Text Chat Completions
GET /v1/models and POST /v1/chat/completions, JSON or SSE and token usage. This is a compatible subset: no Responses, tools, vision, embeddings or Anthropic API.
Multiple clients, one generation
Up to 32 revocable client keys, a bounded queue, fair scheduling, rate limits and deadlines. Inference remains one generation at a time: no continuous batching or multiple resident models.
Memory and model lifetime
Hardware/Metal-aware budgets, configurable ceilings and idle unloading or model retention. More RAM is not a promise that every model will fit.
Control and observability
Queue status, tokens, latency and owner-only Prometheus metrics. The background service needs macOS approval; the Mac must remain powered on and connected.
Lapis Link uses text on a private IPv4 LAN. No photos, documents or voice transfer, and no WAN or Tailscale connections. Its authenticated message encryption is separate from the conventional HTTP API.