LOCAL-FIRST ASSISTANT

Conversation

AI not loaded
Control dashboard →
Account workspace and sync

A separate account workspace for notes, reminders and drafts. Cloud records are readable by the server. Device notes are not uploaded automatically. Chats are session-only in account mode; vault secrets and repeating jobs never sync. Sign in from Account & sync in the control dashboard first. Sync needs opt-in on each page.

Sync off. Nothing uploaded.

CONTROL DASHBOARD

Your workspace.

Real device state. No cloud data has been uploaded.

Local workspace

Loading device state…

Open notes →
Cloud sync

Off. Accounts and sync aren't available yet.

Private vault

Device-only encrypted storage. Unlock inside the app.

Open vault →
Local model

No model loaded on this page.

Open conversation →
Backups

Export ordinary data and the encrypted vault separately.

ACCOUNT FOUNDATION · IN DEVELOPMENT

Guest mode is active.

No account has been created. Cloud sync is off. Your current data has not been uploaded.

Accounts and sync are not available yet. Sign-in will appear only after the real provider is configured and tested.

Planned design

Ordinary notes, reminders and drafts will sync only after you opt in. Chats and vault secrets remain local. Server-readable sync will be labelled clearly.

Model not loaded. Commands still work.Qwen2.5 · on-device

Ask a question, or type /help.

/remember
Save a note
/recall
Find a saved note
/remind
Set a local reminder
/search
Search Wikipedia
/read
Open a text draft
/write
Save a text draft
Commands work without loading AI.

Model answers can be wrong. No messages, purchases or posts are made for you. /search sends your query to Wikipedia.

Saved notes

Only notes you choose to save. No automatic memory guessing.

Reminders appear while this page is open. Missed reminders appear when you return.

Laptop asleep or off? Nothing runs. This is not a 24/7 cloud worker.

Your reminders

Repeating scheduled notes

A deterministic note, not an unattended AI job. At most one overdue note is shown after downtime.

Text drafts

Device-local text drafts. No access to other files on your computer.

Passwords and keys, kept separate from your AI. Nothing from this vault enters a prompt or a network request.

Use at least 12 characters. Forgotten password means no recovery.

Backups and limits

Technical details

AES-256-GCM encryption with a master-password key derived using PBKDF2-SHA256, 600,000 rounds. A strong unique password matters. Forgotten password means no recovery. Auto-locks after 5 minutes without vault activity or when this tab is hidden.

Device-local only, no cloud sync. General app backups do not include the vault. Export its encrypted backup separately. To import, lock the vault and enter the backup's master password above first.

Not a security-audited password manager. Encryption protects stored data, not an infected device, malicious browser extensions, or hostile code running on this website. Unlocked secrets are briefly held in memory. Same-origin hosting integrity matters. Do not store high-risk credentials unless you accept these limits.

Export, import and reset

Browser storage can be evicted. Export backups. Import replaces device-local data only after confirmation. Local companion data is in ~/.pocket-worker.

Qwen2.5 1.5B runs on your device, with optional 0.5B fallback through WebAssembly. No API key, credit card, account automation or paid fallback.

AI modelQwen2.5 1.5B · Q4_K_M (0.5B optional)
Model download1,117,320,736 bytes (small: 491,400,032)
Thinking qualityBasic drafts and simple questions
Background workPage open; local companion for reminder state
Memory privacyDevice-local, not encrypted by this app
Costs outside this appYour internet, power and device
Not a frontier AI or an autonomous employee. No email, calendar, WhatsApp, account login, shell, purchases or unrestricted web browsing. Memory does not sync between devices.

What leaves your device

Chat prompts, notes and drafts stay in this browser or the local companion; sign-in alone does not upload them. Opening the site and downloading the model share normal request metadata with GitHub, jsDelivr and Hugging Face. Wikipedia search sends your search query to Wikipedia. There are no cloud AI calls or app-added analytics trackers. Optional sign-in stores account session tokens in this browser.

Search and sources

/search topic queries Wikipedia only, with source links. It is not a general web search or a completeness guarantee. Retrieved text is untrusted data and cannot run actions.

Cloud AI

Not active. A compatible provider and use case must be verified before cloud AI can be enabled. No API-key input, stored cloud key, or cloud-model requests in this version. Local AI remains the engine.

Two local models

The default Qwen2.5 1.5B Q4_K_M download is 1.117 GB. It is more capable than 0.5B on published benchmarks, but still a small, error-prone model, not a frontier assistant. The 491 MB model remains available from its own button if the larger model fails or is too slow. Loading the fallback is a separate download. Keeping both uses about 1.61GB for weights alone, plus app/runtime files. The app uses a conservative short-context limit. Long messages or notes can be rejected; clear the conversation or shorten them. Accuracy is limited: an arithmetic follow-up failed during testing. Loading can use several GB of RAM; weaker devices may not run either model. Multiple turns, reload with saved history, stopping generation and model switching passed on the hosted origin. The larger model completed real hosted-origin inference in Chrome reporting an 8GB device-memory class, but crashed in a roughly 2GB test environment. CPU first-token delay can exceed a minute. Use the smaller model on low-memory devices; a renderer crash may require reopening the page.

Platforms and what is tested

The web page is designed for browsers on Windows, macOS 11+, Linux, Android 10+ and iOS 17+. Use an up-to-date browser available for your OS. Notes, reminders and drafts do not need the AI model.

Web app actually testedLinux amd64 + Chrome only
Local AI requirementWebAssembly SIMD, modern Chrome recommended, enough free RAM
Phone layout checks320px and 390px browser viewports, not real phones
Other browsers and devicesUntested. AI may fail to load or run out of memory.

Both models use more RAM than its download size. Browser storage may be evicted. Android, iOS and Safari AI inference has not been tested. Automatic external compatibility runtimes are disabled; unsupported browsers can fail to load. This is not a promise that every phone can run the model.

Optional desktop companion

Windows 10+ amd64, Linux amd64/arm64 and macOS Intel/Apple Silicon builds are in the setup ZIP. It records reminders while the page is closed only while the companion is running and the computer is awake. No native phone app or phone companion; phone reminders need the page open.

Linux amd64Binary, launcher and API smoke-tested
Linux arm64 / Windows amd64Cross-compiled; not executed
macOS 12+ Intel / Apple SiliconGo 1.26.6 builds; not executed on Mac
macOS 11 compatibilityOptional Go 1.24.13 builds; not executed
macOS 11 requires an older runtime no longer receiving security fixes. The launcher warns and asks you to type yes. Prefer the browser app or upgrade to macOS 12+. All Mac binaries are unsigned; do not disable Gatekeeper. Windows execution policy may block its installer; no policy is bypassed.

The larger-model install is no longer under 1GB. Download size is not a RAM cap. No cloud sync, account access or unattended AI is added by any platform build.

Privacy and terms

Privacy policy · Terms and disclaimer. These pages explain the app; they do not protect you from every legal claim.

Open software and source notes

Qwen model, Apache 2.0 · wllama, MIT. Updated October 12, 2026. No proprietary prompts or private conversation history are included.