Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add TalissonVitorino/kmp-ios-skills --skill apple-on-device-aigit clone --depth 1 https://github.com/TalissonVitorino/kmp-ios-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/talissonvitorino/kmp-ios-skills/apple-on-device-ai)<a href="https://agentmods.dev/skills/talissonvitorino/kmp-ios-skills/apple-on-device-ai"><img src="https://agentmods.dev/badge/skills/talissonvitorino/kmp-ios-skills/apple-on-device-ai/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/talissonvitorino/kmp-ios-skills/apple-on-device-ai"><img src="https://agentmods.dev/badge/skills/talissonvitorino/kmp-ios-skills/apple-on-device-ai.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00188 | $0.03972 |
| Opus 5 | $0.00094 | $0.01986 |
| Sonnet 5 | $0.00038 | $0.00794 |
| Haiku 4.5 | $0.00019 | $0.00397 |
Grade A, and why
apple-on-device-ai scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 307 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Apple on-device AI (iOS 26+)
Private, no-network generative AI on iPhone/iPad/Mac. This skill routes between the four local runtimes and covers the Foundation Models framework in depth.
- Core ML model loading/prediction, compute units, MLTensor,
.mlpackage→coreml. - OCR, barcode/face/text detection, image segmentation →
vision-framework. - Tokenization, POS/NER tagging, sentiment, word embeddings, on-device translation →
natural-language. - Expose app actions/entities to Siri, Spotlight, and Apple Intelligence →
app-intents. - Fintech note: the system model runs fully on-device — no prompt or user data leaves the phone, which is what makes it usable for balances/transactions. Still treat model output as untrusted (see Safety).
Contents
- Choosing a runtime
- Foundation Models: availability
- Sessions, respond & stream
- Guided generation (@Generable / @Guide)
- Tool calling
- Generation options
- Safety, guardrails & errors
- Context window & performance
- Core ML / MLX Swift / llama.cpp
- Apple Intelligence integration
- Checklist
- References:
references/foundation-models.md,references/alt-runtimes.md
Choosing a runtime
| Runtime | Model | Best for | Cost / caveats |
|---|---|---|---|
| Foundation Models | Apple's built-in ~3B on-device LLM (powers Apple Intelligence) | Chat, summarize, classify, extract, tool-calling — general language tasks with zero model download and guided/structured output | Apple-Intelligence devices only (see availability); not customizable beyond instructions + optional LoRA adapters; text-focused |
| Core ML | Your converted model (.mlpackage) |
Shipping a specific trained model — vision, audio, tabular, or a custom/quantized LLM — with Neural Engine acceleration | You convert & bundle the model; you own updates/size. See coreml. |
| MLX Swift | Any MLX/HF LLM you load | Running larger or newer open LLMs (Llama/Qwen/Mistral/Phi) on Apple Silicon, research, fine-tuning, custom sampling | GPU-heavy; large downloads; iPhone RAM limits model size; you manage weights |
| llama.cpp | GGUF quantized models | Max portability / a mature C++ stack, aggressive quantization (Q4/Q5), reuse across platforms | C/Obj-C++ bridging; you manage the binary, weights, and Metal build |
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 307 lines · 188 tokens per session scan A 3353569534a1
apple-on-device-ai is a skill published in the GitHub repository TalissonVitorino/kmp-ios-skills (12 stars, last pushed 18d ago), licensed MIT. It adds 188 tokens to every session and 3,972 once invoked, about $0.0009 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
compose-multiplatform
Use when building one shared Compose UI in Kotlin across Android, iOS, and desktop — commonMain @Composables, expect/actual, source-set placement, native interop, multiplatform ViewModel/navigation/Koin. NOT a single-platform native build (that is kotlin-android / swift-ios), and NOT Dart/Flutter cross-platform UI…
cmp-new
Scaffold a new MOBILE app — Android + iOS from one Kotlin/Compose Multiplatform codebase — from a bare "create a mobile app" to a green, verified build. Guardrails first: if the user already chose a different framework (React Native, Expo, Flutter, SwiftUI, native), do NOT redirect them here; if they only asked a…
cmp-firebase-connect
Wire a freshly scaffolded CMP/KMP app to its OWN real Firebase project — the #1 post-scaffold manual step. Use this when the user wants to connect their app to Firebase, or asks "connect my app to firebase", "set up google-services.json", "wire firebase", "create a firebase project for this app", "replace the…
add-feature
Add a new conforming vertical-slice feature (Screen + ViewModel + UseCase + Repository + spec + tests + golden tree + nav route + DI wiring) to this Compose Multiplatform app, cloned deterministically from the project's configured exemplar feature (qa/approvals.json's exemplarFeature — home by default). Use this when…
cmp-audit
Adversarial audit of one subsystem of a Kotlin/Compose Multiplatform app against its spec AND against platform semantics — the class of defect desktop-tier tests cannot see (alarms, notifications, PendingIntents, reboot, process death, DST). Use this when the user says "audit the notifications", "double check X for…
add-screen
Add a presentation-only slice — Screen + ViewModel + tests + golden tree + spec + nav route — for an EXISTING entity/repository in this Compose Multiplatform app, cloned deterministically from the project's configured exemplar feature's presentation layer (qa/approvals.json's exemplarFeature — home by default).…