🤖 AndroidClaw

Android-native local AI assistant host — your pocket-sized AI command center.

Inspired by NanoClaw and OpenClaw, built from the ground up for Android.

📖 Project Page · ⬇️ Latest Release · ⚙️ CI / Actions

---

💡 What is AndroidClaw?

Last verified against repo: 2026-03-18

AndroidClaw is a single-APK AI assistant host that runs entirely on your Android device. It turns your phone into a capable AI workstation with durable chat history, extensible tool use, skill-based customization, and scheduled automations — no cloud dependency required.

🎯 Four Core Contracts

Contract	Description
💬 Chat Sessions	Durable history, streaming output, cancel & retry
🔧 Typed Tools	Clear capability reporting, persisted tool-call runtime loop
📜 SKILL.md Skills	Import, enable/disable, precedence, and /slash invocation
⏰ Scheduled Automations	once · interval · cron, with MAIN_SESSION and ISOLATED_SESSION modes

📌 Current Status

AndroidClaw already ships a usable local-first runtime:

• durable chat sessions and persisted message history
• streaming assistant output with cancel / retry
• typed tool calling with persisted tool-call and tool-result turns
• bundled, local, and workspace skills with precedence and enable / disable
• scheduled automations with once, interval, cron, exact-alarm degradation, and task-run history
• a public GitHub Pages project site and installable release artifacts

✨ Features

Chat & Provider

• Durable chat sessions with a deterministic FakeProvider for local testing
• Real-provider support for:
  • OpenAI-compatible
  • MiniMax
  • GLM
  • Kimi
  • Claude via a native Anthropic Messages provider
  • Gemini via its OpenAI-compatible endpoint
• SSE streaming with safe batch fallback on compatible providers
• Visible streamed assistant output, cancel, retry, and clearer failure states
• Budgeted context assembly via ContextWindowManager — no more naive fixed-count slicing
• Per-provider saved base URL, model ID, timeout, and encrypted API key storage in Settings

Skills & Tools

• Bundled, local, and workspace SKILL.md skill loading
• Skill import, enable/disable, precedence ordering, and slash invocation
• Typed tools with a persisted tool-call runtime loop

Automation & Scheduling

• once / interval / cron task scheduling
• Task run history, exact-alarm degradation, and scheduler diagnostics
• Scheduled turns use a durable non-streaming execution path

Build & CI

• Installable qa APK lane + release AAB packaging
• GitHub Actions CI: fast (ktlint + assemble + test + lint) and packaging (APK + AAB)
• Windows-emulator and Android instrumentation validation paths

🏗️ Tech Stack

• Language: Kotlin
• UI: Jetpack Compose
• Persistence: Room
• Background: WorkManager
• Serialization: kotlinx-serialization
• Networking: OkHttp

🚀 Getting Started

Prerequisites

• JDK 17+
• Android SDK (API 31+)

Build

export JAVA_HOME=/path/to/jdk17

# Debug build
./gradlew :app:assembleDebug

# Formatting, unit tests, and lint
./scripts/run_ktlint.sh
./gradlew :app:testDebugUnitTest
./gradlew :app:lintDebug

Release Artifacts

./gradlew :app:assembleRelease
./gradlew :app:bundleRelease

# Or all packaging lanes at once
./gradlew :app:assembleQa :app:assembleRelease :app:bundleRelease

📱 Device Validation (Windows AVD from WSL)

ANDROIDCLAW_JAVA_HOME=/path/to/jdk17 \
  ./scripts/run_windows_android_test.sh \
  --avd AndroidClawApi34 \
  --test-class ai.androidclaw.app.MainActivitySmokeTest,ai.androidclaw.runtime.scheduler.TaskExecutionWorkerSmokeTest,ai.androidclaw.app.OnboardingProviderSetupSmokeTest,ai.androidclaw.feature.tasks.TasksScreenSmokeTest

ANDROIDCLAW_JAVA_HOME=/path/to/jdk17 \
  ./scripts/run_exact_alarm_regression.sh \
  --api34-avd AndroidClawApi34 \
  --api31-avd AndroidClawApi31

📐 Architecture Notes

• Streaming is additive. Providers that don't support streaming still work through the non-streaming path — no breaking changes.
• Ephemeral partial text. Streamed assistant tokens stay ephemeral until the final message is committed to Room.
• Budgeted context selection. The latest important turns and tool-call closures stay within a bounded prompt budget, replacing naive recent-message slicing.
• Hybrid provider strategy. Claude uses a native Anthropic transport; MiniMax / GLM / Kimi / Gemini reuse the OpenAI-compatible runtime path to keep the app small.

🗺️ Roadmap

Status	Item
✅	Native Anthropic provider support
✅	Named provider presets for MiniMax / GLM / Kimi / Gemini
✅	Session-summary generation and context-budgeted prompt assembly
✅	Chat export / share / search
✅	Room kapt → ksp migration
🔲	Baseline Profiles for startup optimization

🚧 Non-Goals (v0)

AndroidClaw v0 intentionally does not include: browser automation, external chat-channel integrations, remote bridge mode, shell execution, or cloud sync.

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to open an issue or submit a pull request.

📄 License

Apache-2.0. See LICENSE.

---

Made with ❤️ for the Android AI community