# MyAi

**Your AI. Actually yours.** Built by Elpmis Software.

The product is now **MyAi** (0.2.2). The former HOLD name remains in compatibility identifiers and historical release evidence. Existing storage, export formats, pairing codes and application IDs are retained. See [rename release notes](/reference/docs/MYAI_RENAME.md).

MyAi is a local AI application for chat, photos, voice, projects and personal agents. Your model runs on your hardware. Ordinary local use requires no account, subscription, API key, wallet or cloud inference service.

Current delivery: MyAi 0.2.3 adds product polish to the free chat, coding and local-agent core. Windows and Android have updated local builds; the website has responsive fixes and a current app capture. See [product polish](/reference/docs/PRODUCT_POLISH.md) for the changes and verification limits, and [free AI scope](/reference/docs/FREE_AI_SCOPE.md) for device capabilities. Apple native validation still requires Apple hardware.

Version 0.2.0 includes a Windows desktop application, an Android native application and a deployable static website. It adds local semantic search, scoped agent delegation, durable edit recovery, Git review, isolated Node tests, private phone pairing and explicitly approved remote connections. This repository records what was actually built and tested in [verification](/reference/docs/VERIFICATION.md) and the [requirements matrix](/reference/docs/REQUIREMENTS.md). Local artifacts are unsigned/development-signed; signing, publication and native Apple-platform validation remain separate gates.

## Desktop

Use Windows x64 with Node.js 24+, npm and room for the runtime and selected models. The recorded development machine has a Ryzen 9 5950X, 32 GB RAM and an RTX 4080. CPU, Vulkan and NVIDIA CUDA were exercised on that machine. Other hardware needs its own validation.

```sh
npm ci
npm run runtime:setup
npm run build
npm start
```

In **Models**, choose a model that fits your device, download it, then choose **Use model**. Downloads show their sizes and licenses and are verified before loading. Smaller models use fewer resources but have limited reasoning and tool reliability. The catalog separates general, coding, reasoning, vision and experimental options; see [actual model evaluations](/reference/docs/models.md).

`npm run runtime:vulkan` and `npm run runtime:cuda` install optional pinned acceleration runtimes. Choose an available runtime in Settings → Device. Failed GPU offload is reported; MyAi does not silently mislabel CPU execution as GPU. macOS and Linux runtime acquisition/bundle paths are implemented, with native execution limitations recorded in [inference](/reference/docs/inference.md).

Chat supports streaming, Markdown/code, editable questions, regeneration, branches, model comparison, per-chat preferences, attachments and export. Search history by text or explicitly index saved content to search by meaning. Windows dictation/read-aloud use installed local voices. Photo questions require a model marked **Reads photos**.

Projects contain instructions, scoped notes, file previews, local retrieval, edit history, Git review and isolated tests. Create an agent with explicit project/tool grants; it can propose changes and bounded subtasks. Every write, external tool call and test execution requires approval. Test execution requires a configured Docker Linux engine and the pinned runtime image; it never runs generated commands directly on the host. See [isolated execution](/reference/docs/isolated-execution-v2.md), [agents](/reference/docs/agent-execution-v2.md) and [security](/reference/docs/security.md).

## Phone and optional connections

The Android app performs genuine local inference after model download, including offline photo and voice paths where supported. It uses native camera/photo/share/QR capabilities. Android emulator results do not certify a physical phone's speed, thermals, battery life or iOS behavior. See [mobile build and acceptance notes](/reference/docs/mobile.md).

Devices enables explicit private-network pairing between your phone and computer. Models and resource allowances are scoped per phone; approved desktop-agent tasks require separate grants and desktop confirmation. Sharing starts off. Optional Power connections send only requests you review to the endpoint you choose. Paid rental and public marketplace features remain gated by enforceable provider controls and owner approval. See [pairing](/reference/docs/pairing.md) and [Power/Network](/reference/docs/power-network.md).

## Build and verify

```sh
npm test
npm run build
node scripts/test-workspace-v2.mjs
npm run licenses
npm run package:win
```

The Windows package defaults to CPU and Vulkan; set `HOLD_BUNDLE_BACKENDS=cpu,vulkan,cuda` to include all three installed, hash-verified runtimes. Packaging validates the exact native runtime manifest. The final installed-app check is `scripts/verify-windows-install-v2.ps1`; it installs into a disposable workspace directory, exercises real inference and services, and uninstalls without deleting its test profile. It refuses to overwrite an existing registered HOLD installation.

Additional scripts exercise real inference, vision, speech, retrieval, agents, Git, isolation and native mobile behavior. Some need downloaded models or local runtime services. Read their scoped reports before interpreting results. Unit tests, native acceptance, model-quality evaluations and hardware observations are different kinds of evidence.

The website builds with `npm run build` and tests with `npm test` in `apps/website`. It uses local assets and no analytics. Public hosting and real download URLs require owner authorization.

## Source map

- `apps/desktop`: sandboxed Electron application and React interface.
- `apps/mobile`: React Native application and native device modules.
- `apps/website`: static website, documentation and deployment output.
- `packages/core`: inference, verified downloads, encrypted storage, project tools and agent orchestration.
- `packages/pairing`, `packages/retrieval`, `packages/git`, `packages/execution`, `packages/power`: bounded capabilities with separate tests.
- `tests`, `scripts`, `.github/workflows`: local verification and build automation.
- `docs`: research, architecture, design, requirements and evidence.

Original MyAi code is Apache-2.0; third-party code, native libraries and models retain their own licenses. Read [PRIVACY.md](/reference/PRIVACY.md), [THIRD_PARTY_NOTICES.md](/reference/THIRD_PARTY_NOTICES.md) and [OWNER_SETUP.md](/reference/OWNER_SETUP.md). No public release or publication is implied by a local artifact or checksum.