MobileClaw
An Android AI agent lab that seats multiple models in one game room to talk, bluff, and vote.
- Source repo
- eggbrid2/mobileClaw
- Stars
- ★ 380
- Last updated
- 4mo ago
- License
- NOASSERTION
- Primary language
- Kotlin
- FA score
- 38/100 · Major gaps
At a glance
- Works with
- Portable with changes
- You'll need
- Typical use
- Model evaluators who want to watch which LLM reasons, bluffs, and leads votes best in one social game.
- Main limitation
- Requires an Android 11+ device plus Accessibility Service and VPN permissions, a heavier bar than a CLI tool.
- Maintenance
- Commits in the last 6 monthslast commit 2026-07-10latest release 0.5.0 (2026-07-03)
- Source review
- 38/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
MobileClaw is an open Android AI agent runtime built with Kotlin 2.2 and Jetpack Compose, shipped as a debug APK for Android 11+ devices. This release centers on an AI arena: you create a room, seat several AI roles, give each role its own configured gateway and model, and a system host runs the flow — calling speakers, collecting votes, announcing tallies, eliminating the highest-vote seat, locking out eliminated seats, and keeping hidden night/event actions from leaking. Around that core sit phone-control agents, role management and memory, local and cloud model routing, MiniAPPs, native AI pages, skill tools, VPN helpers, and a desktop Codex bridge. Builds go through scripts/assemble_debug.sh, releases through scripts/pgyer_release.py, and the README states MIT licensing. It is an on-device application, not a hosted API service.
After a group room is created, the system host drives round progression: it names a role to speak, gathers votes, records hidden actions, and announces results. Each seat can bind a different user-configured gateway and model, so distinct models compete at the same table. Beyond the game loop, the runtime provides Accessibility Service-driven phone control and app automation, VLM-based screen reading, skill routing, MiniAPP execution (with a Javet/Node runtime plan in docs), native AI page transitions, and Mihomo VPN workflows. On the build side, the Gradle task assembleDebug produces app/build/outputs/apk/debug/app-debug.apk; on the release side, pgyer_release.py's build-upload subcommand uploads a build and triggers native in-app update prompts with an APK install handoff.
- Model evaluators who want to watch which LLM reasons, bluffs, and leads votes best in one social game.
- Researchers running model-vs-model experiments by seating differently configured models in the same room.
- Developers building debate rooms, traitor-hunt matches, or custom elimination rules on Android.
- Automation practitioners who need an agent to read the phone screen via VLM and operate apps through Accessibility Service.
- Teams shipping MiniAPPs on device or turning chat results into native AI pages.
- Users managing Mihomo VPN workflows on an Android device.
How do you install or deploy this agent?
You need an Android 11+ device, a working JDK and Android build toolchain, and Git. Build with the README's script:
git clone https://github.com/eggbrid2/mobileClaw.git
cd mobileClaw
./scripts/assemble_debug.shThe output lands at:
app/build/outputs/apk/debug/app-debug.apkInstall that APK on the phone. The README does not document which model-gateway credentials are required or how to enter them in-app. For self-hosted distribution, the Pgyer helper expects secrets in local.properties, .pgyer.env, or environment variables:
python3 scripts/pgyer_release.py build-upload \
--gradle-task assembleDebug \
--notes "MobileClaw AI arena group play release"There is no prebuilt APK link or store listing in the README, and minimum build-tool versions are unspecified — both are known documentation gaps.
How do you use this agent?
After launching, configure a gateway and model per role as described in the README, then create a group room and assign seats. Once the game starts, the system host calls speakers in turn, runs voting, announces tallies, and eliminates and locks the highest-vote seat. Phone-control features require granting Accessibility Service; VPN-related workflows require Mihomo configuration. The Pgyer flow surfaces a native update prompt and hands off APK install. The README gives no step-by-step first-game walkthrough or credential field reference, so consult docs/quickstart.md and docs/group-chat-creation-design.md for the first run.
What are this agent's strengths and limitations?
- Ships a working multi-model arena where a host flow handles speaking, voting, tallying, seat lockout, and hidden-action protection.
- Per-seat gateway and model binding makes model-vs-model comparison a first-class feature rather than an add-on.
- Controls real Android apps through Accessibility Service plus VLM screen reading, covering on-device automation.
- Build and release path is complete: assemble_debug.sh for the APK, pgyer_release.py for upload and update prompts.
- Kotlin 2.2 and Jetpack Compose stack keeps the source auditable and modifiable.
- Requires an Android 11+ device plus Accessibility Service and VPN permissions, a heavier bar than a CLI tool.
- Depends entirely on user-supplied gateways and API credentials, and the README does not document the required fields or a config example.
- No prebuilt APK or store channel is provided; you must build from source.
- Repository license metadata reads NOASSERTION while the README claims MIT — verify before adopting.
- Release is organized around a single fast-moving version, so interfaces and game boundaries may shift.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| MobileClaw This agent | 38 · Major gaps | — | ★ 380 | 4mo ago | Kotlin | — |
| OpenGUI Mobile Operator | 55 · Major gaps | CLIFree + model costs | ★ 1.8k | 5d ago | HTML | Codex · Claude Code · OpenAI API · Claude API |
| Buddy for Android | 53 · Major gaps | Desktop appFree | ★ 308 | today | Java | Codex · Claude Code |
| kwant | 68 · Some gaps | — | ★ 681 | 1d ago | Rust | Codex · Claude Code · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md explicitly enumerates security-sensitive areas (AccessibilityService, dynamic Python/HTTP skills, shell, local/LAN APIs, WebView bridge, VPN) and sets rules against hidden telemetry, deny-by-default, and log redaction; quickstart also says 'grant only the permissions you need' — a plus. But the repository is a high-privilege Android agent that reads screens, controls apps, runs Python, and manages VPN, and the packet contains no code evidence of runtime permission gating, confirmation dialogs, or scoping. least_privilege and user_confirmation are therefore 'principled but unproven'. data_flow_transparency only appears as a contribution checklist in skill-authoring, not as product-level data-flow documentation. sensitive_data_handling has SECURITY.md redaction rules and the 'use a test key' advice, but no code backing. dependency_security is limited to the library list in LICENSE, with no version pinning, audit, or SBOM. external_effects has no rollback/undo description, and rollback is entirely absent (0). source_attribution has MIT plus third-party attributions, but the publisher is unverified, which is deducted.
README and quickstart agree on build steps (assemble_debug.sh, JDK 21, Python 3.11, Chaquopy), and CI reproduces the same toolchain, so self_consistency has basic coherence; however README frames the release as an 'AI Arena group game', drifting from the described phone-control/VLM/Mihomo VPN product, and no source is in the packet to check implementation. dependency_availability relies only on CI's pinned setup-java/setup-python/NDK, with no Gradle dependency locking or version catalog evidence. failure_messages is limited to the quickstart troubleshooting list, with no code-level error handling evidence.
audience_and_scenarios is fairly clear: README targets AI-arena players, quickstart targets developers and real-device testers, skill-authoring targets skill contributors — distinct audiences. capability_boundaries has only the SECURITY.md sensitive-area list and the quickstart 'avoid payments/account changes' advice, lacking an in-product capability boundary or denylist. trigger_precision has only the SKILL.md description field and the 'on-demand' principle, with no routing/trigger implementation evidence. environment_fit documents Android 11+, real device vs emulator, and OEM ROM differences, but no actual compatibility matrix.
information_architecture is clear: README, quickstart, multiple design-doc links, SECURITY.md, AGENTS.md, LICENSE. install_notes is a strong point with full build/install/troubleshooting steps. naming_stability has no version numbers or API/skill naming conventions. examples_and_faq has only two quickstart example tasks and one skill-authoring example, no FAQ. known_limitations has only scattered ROM/root notes, no consolidated limitations. license is MIT with third-party attributions, but repository metadata says NOASSERTION — an inconsistency, deducted. versioning_changelog has only CI tag releases with generate_release_notes, no CHANGELOG file. maintenance_responsibility has only SECURITY.md's 'default branch' statement and an unverified owner contact; publisher identity is unknown.
output_usability has only README screenshots and example task prompts, with no structured output format or result samples. marginal_value's differentiation is 'multi-agent arena on a phone plus phone control', but no source in the packet proves these capabilities work — they are claims only. cost_benefit requires a real device, AccessibilityService, possibly root/ADB helper, and a model API key — high cost with insufficient benefit evidence.
claim_traceability is weak: README capability claims (VLM screen reading, Mihomo VPN, MiniAPP, Codex bridge) are only partially matched by the SKILL.md bridge endpoint list, with no source or test backing for the rest. cross_source_corroboration exists for build steps across README/quickstart/CI, but product capabilities have no cross-source evidence. fact_inference_separation is fair: SECURITY.md and skill-authoring separate rules from examples, but README's marketing prose is mixed with engineering facts.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- The agent requires AccessibilityService screen reading, app control, Python/Shell execution, and VPN management — a high-privilege runtime; the packet shows no code evidence of runtime permission gating or confirmation dialogs, so audit permission requests and scoping before deployment.
- README frames the release as an 'AI Arena group game', drifting from the repository description's phone control, VLM screen reading, and Mihomo VPN; no source is in the packet to verify, so capability claims need independent real-device validation.
- Repository license metadata is NOASSERTION while the LICENSE file is MIT — an inconsistency; publisher identity is unverified and must not be used to infer maintenance quality or safety.
- No CHANGELOG, no version numbers, no dependency locking or SBOM — dependency security and reproducible-build evidence are insufficient.
- No rollback/undo mechanism is described; for phone-control operations, how to recover from a mistaken action is not documented.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Full Changelog: https://github.com/eggbrid2/mobileClaw/compare/0.4.15...0.5.0View on GitHub ↗ -
Release notes excerpt
Full Changelog: https://github.com/eggbrid2/mobileClaw/compare/0.4.15...v0.5.0View on GitHub ↗ -
Release notes excerpt
Full Changelog: https://github.com/eggbrid2/mobileClaw/compare/0.4.14...0.4.15View on GitHub ↗
FAQ
Is MobileClaw open source?
Is MobileClaw safe to use?
Is MobileClaw still maintained?
Who makes MobileClaw?
Thanks — we use this to improve the page.