Artfull AI0.0.1 build 13 · development status
This catalogue explains what each Artfull AI capability does, which tier it belongs to and how far its validation has progressed. Build 13 has passed iPhone and iPad simulator checks and a signed iPhoneOS build, but Artfull AI is not available on the App Store.
Available in the development build
Current capabilities use genuine development captures where a relevant screen is available. A screenshot may illustrate more than one closely related capability. An unlabeled feature works with no known reported issue; a status label appears only where implementation or testing work remains.
Connect directly to an Ollama server, discover its models and keep model selection scoped to that connection. The app never silently reroutes an offline choice to a cloud provider.
Use LM Studio through its OpenAI-compatible local API with the same direct-server boundaries as Ollama. Controlled API tests pass, and direct connectivity was confirmed on a physical iPhone in 0.0.1 build 7. Broader release validation remains open.
Choose Apple’s on-device intelligence when the device and operating system make it available. Apple Intelligence chat has been physically confirmed on an iPhone 18 Pro Max. This route is local-only and does not silently fall back to a network model. Wider device and capability compatibility still depends on Apple’s hardware and operating-system support and has not all been tested.
Browse a filtered MLX model catalogue, remove files and run compatible installed models offline. Model rows now offer Download, More Information and Cancel. Model download and local chat inference have been physically confirmed on an iPhone 18 Pro Max, but manual interaction with the new row prompts has not yet been confirmed. Downloads require an explicit internet action. Broader compatibility still depends on iOS or iPadOS 17 or later, available RAM and storage, model size and quantisation, and thermal conditions; not every model or supported device has been tested.
Connect to Ollama Cloud through a fixed HTTPS origin and a connection-scoped Keychain credential. Provider access and charges remain separate from the app purchase.
Use a scoped developer API key, discover available models dynamically and retain usage metadata when the provider supplies it.
Configure a compatible gateway and token, detect supported endpoints and show a clear failure when its gateway API or model catalogue is unavailable.
Sign in to a separately maintained Router, use its server-supplied instances and models and access Router-managed history across devices without mixing it into direct-server history.
Configure temperature, context and response length, top-p/top-k, repeat penalty, system prompt, timeout and keep-alive for the relevant server profile. Basic uses the standard response settings.
Manage models without leaving Artfull AI. Ollama provides installed and running inventory, cancellable downloads, loading, unloading and deletion. LM Studio 0.4.0 or later provides its native v1 inventory, cancellable downloads, loading and unloading; models must be deleted in LM Studio because its API does not expose deletion.
These lifecycle controls passed the iPhone and iPad simulator harnesses and the full build. The signed build 12 app was also installed and launched on a physical iPhone; that installation is not a claim that every lifecycle action was manually exercised on the physical device.
Share a downloaded On-device MLX model with a nearby Artfull AI device. The connection is encrypted and requires a six-digit pairing code plus approval on the sending device. Bounded manifests and SHA-256 validation protect the receiving installation.
iPhone and iPad simulator checks, the full simulator build, signed iPhoneOS build, installation and screenshot checks pass. A transfer between two physical devices, an interrupted transfer and a wrong-code attempt remain release gates.
StoreKit 2 entitlement checks, Restore Purchases and transaction updates are implemented. Deterministic iPhone and iPad simulator tests cover Basic, purchased, refund or revocation, restore eligibility and the verified-entitlement policy. The local StoreKit fixture is excluded from Release builds.
Final price, App Store Connect product configuration and live purchase testing remain release gates.
A SwiftUI interface designed for phone and tablet layouts, with system, light and dark appearance and native sharing, files, photos, camera and speech surfaces.
Save server names, addresses and connection-specific settings away from the composer. Basic keeps one Ollama or LM Studio profile; Advanced removes Artfull AI’s fixed saved-connection limit.
Messages in the open conversation remain available as context while you chat. Basic does not reopen saved conversations after the current session.
New user and assistant messages show a time formatted for the device locale. Older messages show an abbreviated date with the time. Timestamps persist through saved history, continuation, branching and Advanced backup export/import; legacy messages without a timestamp remain readable and are not assigned a fabricated time.
Each response keeps its model and provider label even after the selected model changes, so the origin of an answer stays visible.
Display finished Markdown-style responses with headings, lists, tables and code. Optional streaming remains available.
Copy a finished response or share its formatted output through the native system share sheet.
Continue an interrupted or length-limited direct-server response in place without duplicating text that has already been generated.
Export complete direct-server conversations as bounded, versioned JSON backups with server provenance. Imports add validated conversations to the selected direct server without replacing existing history or moving credentials.
Authorize an exact tool for one message at a time. Build 12 exposes only bounded on-device arithmetic and current date and time, allows at most three calls in a response and keeps every tool activity visible in the conversation. It exposes no files, URLs, shell commands or arbitrary code.
Web Search is an Advanced feature for direct AI Model Server chats. Enable it for one message at a time when the model needs current reference information.
You need your own Brave Search API account, active plan and API key. The key is stored in the device Keychain. Only the search query you type is sent to Brave Search; attachments and chat history are never included.
Search results are bounded, treated as untrusted reference text and shown as source links with the answer. Brave pricing and usage limits apply. See Brave Search API information and sign-up.
This feature is not currently released. A successful live keyed Brave Search smoke test remains an acceptance requirement.
Retain, reopen, search, rename, pin and archive conversations. Direct and on-device histories remain separated by connection; Custom LLM Router history is managed by its server.
Create a new conversation branch from any message while preserving the original conversation unchanged. The branch copies history only through the selected message and keeps attachment storage bounded.
Stop generation, keep partial output and retry explicit failures without creating an endless polling loop. Wider physical-device background testing remains a release check.
Use native dictation where language and hardware support it.
Artfull AI keeps normal device auto-lock enabled rather than forcing the screen to stay awake.
View response metadata such as provider, model, timing and token counts when available. The local journal is designed not to store prompt, response or document text.
Attach images selected from the photo library or captured with the camera. Image attachments remain part of Basic as well as Advanced.
Extract text from PDF, scanned PDF OCR, Word, Excel, PowerPoint, plain text and CSV files within documented size and page limits. Old binary Office formats require conversion.
* Artfull AI does not impose a fixed saved-connection limit. The practical number depends on available device storage and system resources.
Future work
These features are not available now. Their Basic or Advanced tier is “To be decided” unless a later product decision explicitly assigns it.
Work across larger documents with grounded page, sheet and slide references and on-device chunking where practical.
Ask two models on the same server and keep each answer, provider label and statistics separate.
Save repeatable instructions and generation settings without silently changing the active connection.
Offer optional next questions while identifying where they are processed and any provider cost they may create.
Add optional retention and auto-delete controls, explicit clear-history confirmation and device-authentication protection.
Add secret-free diagnostic export, connection latency and status, adjustable text, Dynamic Type, optional haptics and iPad keyboard shortcuts.
Improve tables, syntax highlighting and maths, then add read-aloud and a more conversational voice interface.