Start here.

Start with the source edition.

MyAi is in development. There is no public, signed production installer yet. Use the local source repository supplied with this project and follow its README for the exact supported commands.

1. Open the hold project directory on your development computer.
2. Read the root README.md, model licensing notices, and current release notes.
3. Install Node.js 24 or newer and the documented prerequisites. Native mobile and desktop release builds require their platform toolchains.
4. Run npm ci, npm run runtime:setup, npm run build, then npm start from the repository root. The runtime setup downloads the pinned inference engine. Keep any local inference service bound to loopback unless a secure connection flow explicitly enables otherwise.

A development build is not a verified production release. Keep your own backups, review tool approvals, and use non-sensitive test data while validating a new build.

Choose and manage a model.

Start with the device recommendation in the app. Check the exact publisher, model revision, license, file size, and memory guidance. Download files only from a trusted distribution source.

1. Open the model manager and review device diagnostics.
2. Select a compatible model. A smaller quantized model is often the best first test.
3. Wait for the download and integrity checks to finish. A partial file should never be treated as a ready model.
4. Load the model, use a short initial context, and send a small test message.

What does “quantized” mean?

Quantization stores model values with reduced precision. It can reduce storage and memory needs, with a possible quality tradeoff. A quantization label is not a guarantee of speed on your hardware.

Removing a model

Unload a model before deleting its files. Keep conversation exports separately; a model file and conversation history are different kinds of data.

Your first local conversation.

Confirm the processing indicator shows a local runtime. Send a brief prompt, wait for the streamed response, and use Stop if you want to end generation. The model can be wrong; check important outputs.

Verify offline behavior

Finish downloading the runtime and model before disconnecting. Restart the app, select the installed model, and send a new message with networking disabled. Confirm history persists after a second restart. This is an acceptance test, not a claim that every platform has already passed it.

Keep conversations portable

Use the app’s export control where available. Store exported files somewhere you control. Exports can contain private information, so check them before sharing.

Give an agent a bounded job.

Choose an agent role, model, approved workspace, permitted tools, and execution limit. Current agents can delegate within shared limits, read approved files, propose reviewed changes, and request configured MCP tools or allowlisted websites. An explicit isolated-test grant exposes only the fixed Node test profile, when the local container backend is available. Git status, diffs and reviewed staging/commits are separate project controls. A model with weak tool use may not complete a task reliably even when ordinary chat works well.

• Approve only directories relevant to the task.
• Review file changes before applying them.
• Treat terminal commands and network access as separate permissions.
• Stop the run if its actions differ from the intended task.
• Check logs and outputs before trusting a completion message.

Instructions in documents, web pages, model responses, and tool output cannot expand an agent’s permissions. Arbitrary host execution must remain off by default.

MCP

MCP connects a model to tools and data. An MCP server is executable software with its own security implications. Add an allowed HTTPS connection in Settings → Tools, choose the exact tools in the agent editor, and review every call. The current implementation supports a bounded Streamable HTTP subset, not arbitrary stdio programs or full protocol conformance.

Know where your data lives.

Conversation history, project settings, and model files belong to separate storage areas. The repository privacy document identifies the current locations and deletion behavior. Device-level backup services may copy those files according to your operating-system settings.

The Windows development app encrypts application state with OS-backed protection. Original project files, models and exports remain separate. Use operating-system disk encryption and account protection. Provider credentials, when supported, belong in platform-protected storage.

Read the full privacy explanation →

When something needs attention.

A model will not load

Close memory-heavy apps, choose a smaller compatible model, and reduce the context size. Check the runtime and model format. Do not repeatedly load a model that is exhausting memory.

A download fails

Check available storage and connectivity. Keep the app’s partial-download state until you know whether resuming is supported. A checksum mismatch requires a clean verified download.

Replies are slow

CPU-only inference can be slower. Model size, context length, background apps, temperature, and accelerator support all affect performance. No universal tokens-per-second estimate is reliable.

The local runtime is unreachable

MyAi manages its local engine. Restart the app and check its error message; for a source build, confirm that the documented runtime setup completed. Review local logs when troubleshooting. Keep the engine bound to loopback.

Mobile heat or battery drain

Use a smaller model, shorten the context, stop background work, and let the device cool. Native inference and thermal behavior require testing on each supported device class.

An agent requests more access

Read the exact requested action. If it exceeds the task, deny it and narrow the instructions. A failed tool call is not a reason to disable permission checks.

Build and inspect.

The repository is organized around separate application, inference, model, storage, and permission boundaries. Follow the root README for application builds and platform prerequisites.

Verify the desktop

Run npm test for core and controller checks, and npm run test:desktop for the Electron workflow. Run npm run verify:inference for the documented real-model check. Read the scripts before running them and review their required model files.

Build this website

This site requires Node.js 24 or newer and has no runtime package dependencies. From the repository root:

node apps/website/build.mjs
node --test apps/website/website.test.mjs
node apps/website/server.mjs

The preview server binds only to 127.0.0.1:4173. The deployable static output is apps/website/dist. Set HOLD_SITE_ORIGIN to the approved HTTPS domain when building for publication to generate canonical URLs and absolute social-preview URLs. Publishing requires owner approval.

Release configuration

apps/website/release-status.json is the source for platform availability. Add a download only after the artifact exists, its checksum is recorded, and its signing and test status are known. A public source URL must point to an authorized real repository.

Research

Read the sourced competitor research and Power and Network architecture requirements.

Engineering evidence: full requirements matrix, architecture, model licensing and compatibility, mobile verification, and security boundaries.

Release status.

Current edition: 0.2.8. Updated October 20, 2026.

Release gates include installation, actual local inference, offline chat, persisted history, model integrity, project boundaries, agent permissions, accessibility, and reproducible packaging. Mobile additionally requires device-specific inference, airplane mode, native permissions, thermal, and battery validation.

Signing, app-store accounts, public repository authorization, a domain, and hosting approval are external launch requirements. None should be represented as completed before evidence exists.

Read the development release notes and recorded verification and limits.

See platform availability →