Jandal AI

Play Store release in progress. Pre-release builds are available from GitHub Releases.

Jandal AI is a local-first Android assistant. It combines on-device chat, long-term memory, deterministic Android actions, voice input/output, and native tool calling while keeping user data on the device by default.

How it works

Jandal is built around a Brain-Memory-Action model:

• Brain - Gemma-4 E-2B / E-4B runs locally through Google AI Edge LiteRT / LiteRT-LM. A lightweight QuickIntentRouter handles deterministic fast paths for common device actions.
• Memory - local Room storage plus sqlite-vec / Arctic Embed M v1.5 support semantic recall, conversation history, core memories, and episodic summaries.
• Action - native Kotlin skills execute Android actions such as alarms, timers, lists, notes, weather, media controls, messages, email, calendar, navigation, Wikipedia, and unit/currency conversion.

Current launch status

Launch readiness is tracked in GitHub rather than duplicated in this README:

• #1014 - Jandal v0.1 Play Store release candidate is the parent launch epic and current launch tracking issue.
• The launch:blocking label and Jandal Launch Backlog project views are the live launch blocker queue and filtered work dashboard.
• docs/PLAN-launch-slice.md is a historical launch-slice snapshot, not the current plan.
• #1255 - Launch Plan: ordered implementation sequence and release gates is the now-completed model-readiness preflight tracker, not the active launch plan.

The current launch gates are:

1. Release-candidate golden journeys (#1329).
2. A higher-quality local TTS option (#1447), with current candidate safety/evidence tracked in #1449 and #1485.
3. Play Store, legal, and policy readiness (#441), including #1260–#1264 and #1474.

Wake-word launch qualification is complete. Broader permission orchestration, evidence-platform, and wake-word optimisation work remains post-launch unless a new release-candidate regression proves otherwise.

Feature areas

Jandal currently includes:

• local chat with markdown rendering, multi-conversation management, streaming generation, and model/runtime status;
• local memory with core memories, episodic summaries, semantic search, and memory management screens;
• deterministic Android skills for lists, notes, alarms, timers, date arithmetic, media controls, unit/currency conversion, weather, messaging/email flows, navigation, and Wikipedia;
• local-first Lists with two-level parent/child groups and device-local expand/collapse state, plus optional bidirectional sync for explicitly selected Lists to a user-configured Nextcloud Tasks server over CalDAV/VTODO; Room-backed local state remains authoritative and usable offline;
• drawer-accessible tools such as Lists, Alarms, Notes, Meal plans, and Settings;
• deterministic meal-planning sessions with plan approval, recipe persistence, favourites, replacement/regeneration, and ingredient export to lists;
• push-to-talk voice input, optional spoken chat replies, per-message speaker playback, configurable TTS voices, and wake-word/default-assistant infrastructure;
• local model management and model availability states for required, optional, preparing, gated, and unavailable model paths.

Future roadmap work includes Dreaming Engine background consolidation, Wasm skill extensibility, broader device optimisation, and richer multimodal/vision workflows. These are intentionally separated from the launch plan unless explicitly promoted.

Tech stack

Models and downloads

Model files are not committed to this repository. Some model downloads are gated and require a Hugging Face account plus acceptance of upstream terms. Arctic Embed M v1.5 is ungated and temporarily hosted at a public rolling GitHub release; the validated artifact's SHA-256 provenance is recorded in docs/research/1559-arctic-phase-b-calibration.md. Long-term publication to litert-community is tracked separately in #1563; GitHub release immutability is not a #1559 gate.

See models/README.md for the current model file reference, approximate sizes, ADB setup notes, and device-specific guidance.

Key launch-relevant examples:

Privacy model

Jandal is designed as a local-first assistant:

• chat, memories, notes, lists, and model execution are local by default;
• production builds should not include telemetry;
• external network exchange occurs only for explicit web-backed skills, configured model/voice downloads, or user-enabled integrations, and is limited to the data needed for that integration;
• explicitly selected Lists may synchronise with a user-configured Nextcloud server while unsynchronised Lists remain local-only;
• Hugging Face sign-in is used only for gated model access.

Documentation

• models/README.md - model file setup, sizes, and ADB paths.
• docs/LEGAL_AND_ATTRIBUTION.md - launch attribution and licence checklist.
• docs/SPECIFICATION.md - broader product and architecture specification.
• docs/ROADMAP.md - broader roadmap notes.
• docs/testing/ - automated test harness, evidence schema, and QA documentation.

Licence and attribution

The source code is licensed under the Apache License 2.0. See LICENSE.

Third-party notices are maintained in NOTICE. Launch readiness for third-party libraries, voice assets, and gated model downloads is tracked in docs/LEGAL_AND_ATTRIBUTION.md.