Building on YeongSil: What the Developer Platform Will Offer

A consumer AI device that ships with the right architecture is also, eventually, a platform. The smartphone took eighteen months from launch to the App Store; the next category of AI hardware will not have that grace period. Developers will need a clear set of primitives, a clear revenue model, and a clear path from prototype to shipping skill on day one.

This post is the most detailed public preview of the YeongSil developer platform: what the SDK exposes, how the Skill Store will work, what the revenue split is, and which categories of skill we think will define the first twelve months.

The platform thesis

YeongSil is a standalone device with three things developers cannot easily replicate: an always-on local context (camera, microphone, presence), a per-user encrypted RAG memory, and an action surface that reaches outside the device (phone bridge, messaging, calendar). A "skill" is a third-party package that uses some or all of these primitives to deliver a focused capability without the developer having to ship hardware or rebuild memory infrastructure.

The closest existing model is Alexa Skills, but with three meaningful differences. First, skills run against a richer context than Alexa ever had — visual, document-aware, durable. Second, the revenue split is structured for skills to be commercial products, not lead-generation for an app. Third, the device is a single, controlled hardware target rather than a sprawling family of speakers with inconsistent capabilities.

The SDK primitives

The YeongSil SDK exposes four capability surfaces. We are publishing a preview specification for early-access partners in late 2026; the shape below is stable.

Memory primitives. A skill can request scoped read access to specific document categories the user has tagged (e.g. "fitness logs," "legal contracts," "language-learning materials"). The user authorises the scope explicitly; the skill never sees documents outside its scope. The skill can also write to its own private namespace, indexed alongside the user's documents but invisible to other skills.

Perception primitives. A skill can request temporary access to a camera frame ("show me the label") or a microphone segment ("listen to the next thirty seconds"). Access is bounded — single frame, single segment, single language-model turn — and surfaced to the user with a hardware indicator while it is happening. This is the same primitive YeongSil itself uses internally; we did not invent a special privileged interface.

Action primitives. A skill can request to make a call, send a message via the phone bridge, set a reminder, or add a calendar event, subject to user authorisation. It cannot escape the action surface — there is no general "run arbitrary code on the user's network" primitive, and there will not be one. The action surface is the same conservative list we ship at launch (see [our post on what makes AI personal](/blog/what-makes-ai-personal) for the rationale).

Language-model primitives. A skill can invoke the on-device LLM (or, with the user's authorisation, the cloud-compute fallback) with a prompt and retrieved context, and receive a structured response. Skills are billed for compute on a metered basis; pricing tiers will be published with the SDK.

The Skill Store

Skills are distributed through a single store, surfaced through voice (the user says "install the fitness coach skill") or through the optional companion app for skills that need a visual setup step. The store enforces three commitments.

Code review before publication. Every skill is reviewed against a published checklist before it can ship to users — scope minimisation, privacy disclosure clarity, action-surface usage. We are not running an open registry. The trade-off in slower onboarding is what makes the platform durable for users.

One-tap install, one-tap uninstall. Skills install in seconds via voice, with the scope disclosure read aloud before install. Uninstall is symmetric — one voice command removes the skill and purges its private namespace, with the user given the option to export the namespace first.

Standard privacy contract. All skills inherit YeongSil's base privacy posture: no training on user data, encrypted at rest, exportable, deletable. Skills cannot opt out of these properties; the SDK does not expose the primitives that would let them. This is the architectural commitment that makes the store browsable without the user having to read seven different privacy policies.

The revenue model

YeongSil takes 30% of net skill revenue. Developers keep 70%. The split applies to one-time purchases, recurring subscriptions, and metered usage. Compute and platform infrastructure costs to the developer are billed separately, transparently, at cost-plus.

This is more generous than the 70/30 incumbents (Apple, Google) and a cleaner structure than Alexa Skills, which historically had no consumer-paid model at all. The bet is that a fair split with a small but engaged installed base produces better skills than a punitive split with a larger one. We are happy to be measured on whether the bet pays off.

Three things are explicitly excluded from the platform fee. Direct enterprise licensing of a skill to a single customer (e.g. a hospital licensing a skill bundle for its clinics) is between the developer and the customer, with no platform cut. Open-source community skills that are free at the point of use are not metered. Skills that ship as part of an organisational deployment (NGOs, accessibility partnerships — see our [accessibility page](/accessibility)) are operated under separate terms.

The categories we expect to define year one

We have spent eighteen months in private conversation with would-be developer partners. Four skill categories surfaced repeatedly. We list them not as a sealed roadmap but as a reflection of where the demand seems concentrated.

Personal coach skills. Fitness, nutrition, meditation, sleep. The common shape is a skill that ingests the user's daily logs (workouts, meals, sleep data from a Garmin or Apple Health export), references a coaching protocol the developer has authored, and produces ambient prompts during the day — "your run window is closing, leave in ten minutes" or "you're behind on water for today." The fit with YeongSil is strong because the value of a coach is in its presence, not its dashboard, and YeongSil is built for presence.

Profession-specific skills. A legal helper for solo practitioners that walks through the standard clauses in a contract upload and flags non-standard language. A clinical helper for small practices that surfaces patient history at the moment of consultation. A construction helper that reads project specifications and answers questions on site. Profession-specific skills benefit most from YeongSil's memory architecture because the relevant documents (case files, patient records, project specs) are exactly what RAG is best at.

Language-learning skills. A conversational tutor for Mandarin, Arabic, German, French — a skill that the user can speak with for fifteen minutes during morning coffee, that remembers the user's vocabulary level and grammar gaps across sessions, and that escalates difficulty as the user improves. This is a skill category that smartphones have failed for years (the friction of opening Duolingo at 7am defeats most learners). Ambient delivery removes the friction.

Local-business assistant skills. Skills built for specific small-business categories: a hotel-management skill, a restaurant-front-of-house skill, a clinic-receptionist skill (see our [small business use cases](/blog/yeongsil-small-business-use-cases)). These are typically built by local developers who know the category deeply and benefit from YeongSil's voice-first, hands-free posture more than they would benefit from another web app.

Early access programme

We are opening early access to twenty developer teams in Q1 2027, ahead of the consumer launch. Early access includes hardware preview units, SDK previews two months ahead of public availability, direct engineering support during integration, and featured placement in the store at launch. Applications open in December 2026.

The selection criteria are concrete. We want skills that exercise at least one of the four primitive surfaces materially (a wrapper around a website does not need YeongSil; a skill that uses durable memory plus voice plus action does). We want developers who can ship — running products, shipped apps, evidence of execution. And we want a mix of geographies, because the small-business and accessibility skill categories are deeply local.

If you are a developer building in this category and want to be on the early-access shortlist, the application form opens at the end of December 2026. In the meantime, the architectural posts that explain what the platform is built on — [RAG vs fine-tuning](/blog/rag-vs-fine-tuning-personal-ai), [what makes AI personal](/blog/what-makes-ai-personal) — will give you a clear picture of the primitives you will be building on.

[Join the waitlist](#waitlist) and select "Developer" to be the first notified when applications open.