Buddy for Android
Let a floating AI chat bubble read your Android screen and operate apps for you.
- Source repo
- ghorbelhamdi/buddy-android
- Stars
- ★ 308
- Last updated
- today
- License
- Apache-2.0
- Primary language
- Java
- FA score
- 53/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Android users who want Claude Code or Codex to handle tasks in messaging, shopping, or smart-home apps.
- Not a fit if
- People who need to automate other apps on an iPhone
- Teams that require every request to stay on the phone
- Website
- ghorbelhamdi.github.io/buddy-android ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v2.3.16 (2026-11-20)+12.1 ★/day over 7 days
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
Buddy is an app for 64-bit Android 11 and newer that launches Claude Code, Codex, OpenCode, Antigravity CLI, or Grok from a floating bubble. Its Alpine Linux environment runs those command-line clients through proot, while Buddy’s accessibility service supplies phone tools for reading the screen, tapping, typing, scrolling, and opening apps. The tools are exposed over a local MCP server bound to 127.0.0.1, and agent replies stream back into the chat. Buddy can ask for approval before sending, posting, buying, or deleting, and also supports notification rules, voice, background operation, and a background browser. It is installed as an APK and requires the user to enable the accessibility service and sign in to the selected agent as needed; agent requests are sent to the relevant provider.
The user enters or speaks a task in Buddy’s bubble, and Buddy hands it to the claude, codex, opencode, agy, or grok client in its Linux environment. The client calls Buddy phone tools through the local MCP server to read the current screen, tap, type, scroll, or open apps; the results return to the chat. Buddy presents approval controls before actions such as sending, posting, buying, or deleting. Users can also configure notification triggers, use voice, operate apps on a background screen or in Buddy’s built-in browser, and enable Linux tools to build Android APKs.
- Android users who want Claude Code or Codex to handle tasks in messaging, shopping, or smart-home apps.
- People who want to review and approve a message, post, purchase, or deletion before it goes through.
- Users who want to draft notification-triggered replies and send them after approval or a stoppable countdown.
- People who want an agent to guide them through phone settings one step at a time, waiting for each tap.
- Developers and advanced users who want to build an Android APK on the phone or browse in the background.
How do you install or deploy this agent?
You need a 64-bit Android 11+ phone. The README provides a direct APK download. On first launch, follow the setup guide to choose an agent, enable the accessibility service, install and sign in to that agent, then try a first task. The README says Google Play declined Buddy for production, so installation is via APK.
How do you use this agent?
Open Buddy’s floating bubble, choose a configured agent, and type or speak a task. The agent can operate the screen through Buddy’s accessibility service and local MCP phone tools; approve, deny, or change the action when prompted before sending, posting, buying, or deleting. Claude Code, Codex, Antigravity, and Grok require their respective account or plan; OpenCode’s free models need no account. Requests are sent to the provider you use.
What are this agent's strengths and limitations?
- Operates existing Android apps through the accessibility service without root, a computer, or adb.
- Offers Claude Code, Codex, OpenCode, Antigravity CLI, and Grok within one app.
- Buddy itself enforces approval before sending, posting, buying, or deleting.
- Includes notification automation, voice, background mode, a background browser, and on-phone APK building.
- Exposes a local MCP server for use by the user’s own agent sessions.
- Requires 64-bit Android 11+; the README says cross-app automation on iPhone is unavailable.
- Requires a powerful accessibility permission; the README identifies prompt injection as the main risk.
- Agents need their respective provider accounts or plans, and requests are sent to those providers.
- The README labels version 2.3 early and says it still has rough edges.
- Some phone brands may require Buddy’s battery use to be set to Unrestricted, and some background features require extra setup.
How does this agent compare with similar options?
The README distinguishes Buddy from iPhone cross-app automation, which it says iOS does not allow, and says Buddy needs no computer, adb, or root. It offers OpenCode’s free models as an account-free agent option; other agents use their respective accounts or plans.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Buddy for Android This agent | 53 · Major gaps | Desktop appFree | ★ 308 | today | Java | Codex · Claude Code |
| Clawd Cursor | 72 · Some gaps | MCP serverFree + model costs | ★ 406 | 1d ago | TypeScript | Codex · Claude Code |
| AIHawk – AI Browser Agent | 52 · Major gaps | MCP serverFree + model costs | ★ 33k | 1d ago | Python | Codex · Claude Code |
| Browsentic | 75 · Some gaps | — | ★ 300 | 15d ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The permission surface is broad: accessibility service, notification access, install packages, battery-optimization exemption, exact alarms. PRIVACY.md explains each purpose, but explanation is not minimization; the accessibility+notification+install combination is disproportionate for an ordinary user, so least_privilege is 1. User confirmation: README/SECURITY/SKILL all state that send, pay and delete stop on an approval card enforced by Buddy rather than the agent, and install_apk requires the user to tap Install, so 2. Data-flow transparency: PRIVACY.md names each provider per agent, voice engines, download sources and the GitHub update check, so 2. Sensitive data: it claims no passwords/codes/payment entry, screen read only during a request, unmatched notifications discarded, but these are documentation assertions without code in the packet, so 2. Dependency security: CI pins the NDK and runs a reproducible-build check, but third-party submodules (proot/talloc/libandroid-shmem) and Alpine downloads have no stated hash/signature verification, so 1. External effects: approval cards, countdowns and stop controls, so 2. Rollback: only over-install preserving data and uninstall deleting it; no per-action undo or backup, so 1. Source attribution: explicitly independent and unaffiliated with Anthropic/OpenAI/Google/xAI, with third-party licenses listed, so 2.
Self-consistency: README, PRIVACY, SECURITY and SKILL agree on approvals, 127.0.0.1, per-install token and no-root claims, so 2. Dependency availability: relies on Alpine mirrors, npm, vendor CLIs and the GitHub API, all mutable external sources with no version pinning or offline fallback described, so 1. Failure messages: SKILL.md gives recovery steps for install failure, timeout and settings redirect, but product-level error handling is not evidenced, so 1.
Audience and scenarios: aimed at ordinary Android users with messaging, shopping, smart home, notification auto-reply, voice and local models, so 2. Capability boundaries: Android 11+/64-bit only, no iOS, local small models unreliable on multi-step tasks, Play rejection stated, so 2. Trigger precision: the SKILL.md description trigger is broad ('make, build, change or install an Android app') and only one skill sample is present, so 1. Environment fit: Termux build path, local Ollama/LM Studio and Tailscale support, so 2.
Information architecture: README is well structured and links GUIDE/PRIVACY/SECURITY/CONTRIBUTING/benchmark, so 2. Install notes: APK download, first-launch guide, Termux build and signing-key env vars are documented, so 2. Naming stability: version lives in BuildInfo.java and README says 'Status: 2.3, early', but there is no changelog or versioning policy, so 1. Examples and FAQ: FAQ, screenshots, video and sample tasks, so 2. Known limitations: early status, tested device, local-model limits and Play rejection are stated, so 2. License: full Apache-2.0 text plus third-party license listing, so 3. Versioning/changelog: only CI-generated release notes; no in-repo CHANGELOG, so 1. Maintenance responsibility: single author, SECURITY.md supports only the latest release, no governance or maintainer list, publisher unverified, so 1.
Output usability: approval cards, folded tool calls, live in-chat viewing and post-install verification make output actionable, so 2. Marginal value: wiring desktop coding agents to phone accessibility actions is a differentiated capability, so 2. Cost-benefit: free but consumes the user's own plan quota; local models need ~24k context and failed multi-step tasks on an 8GB GPU, so value for ordinary users is uncertain, so 1.
Claim traceability: PRIVACY says the source is public and checkable and SECURITY names Guard.java/Reviewer.java/Bubble.java/system-prompt.md, but those files are not in the packet, so 1. Cross-source corroboration: README, PRIVACY and SECURITY agree and CI has fixture tests plus a reproducible-build check, but all are self-reported with no independent verification, so 1. Fact/inference separation: docs distinguish tested devices from expected behavior and flag the early status, so 2.
- The combination of accessibility, notification access, install-packages and battery-optimization exemption is powerful; verify the APK source before installing, and the publisher identity is unverified.
- Approvals are enforced by Buddy itself, but Guard.java/Reviewer.java/Bubble.java are not in this packet, so it cannot be confirmed that every send/pay/delete path is gated.
- Prompt injection is named as the main risk; screen text and notification content are sent to the chosen AI provider, so sensitive data can leave the phone.
- Third-party submodules and Alpine/npm downloads have no stated hash or signature verification; assess supply-chain risk yourself.
- Local small models are unreliable on multi-step tasks and may falsely report success; avoid high-stakes actions such as payments.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v2.3.16 Buddy v2.3.16
Release notes excerpt
Fixes - Models on this phone (#67): chats in lite mode now get only Buddy's phone tools. MCP servers from your own OpenCode settings used to come along. A large one put every request at ~250k tokens, five times the model's window, so even a new chat failed - Live voice (#65): the voice's words in the bar at the bottom start with 🔊 and clear once it's done. They looked like something you said, as if the speech recognition had picked up Buddy's reply Full Changelog: https://github.com/ghorbelhamdi/buddy-android/compare/v2.3.15...v2.3.16 …View on GitHub ↗ -
v2.3.15 Buddy v2.3.15
Release notes excerpt
Models on this phone - Fixed: no model started on phones without Snapdragon (Pixel, Exynos, MediaTek): the engine didn't find its CPU part (#67) - Faster answers: models on this phone now answer without thinking first, as Qwen3.5 2B is made to. A step takes a second or two instead of 20–30 seconds, and a new chat no longer waits on its title. Think before answering (Settings → Agents and models → Models on this phone) turns it back on - "What's on my screen?" in a background chat on a small model reads your screen while the agent's own screen is still empty …View on GitHub ↗ -
v2.3.14 Buddy v2.3.14
Release notes excerpt
Models on this phone (#63) Settings → Agents and models → Models on this phone: download a model and Buddy runs it on the phone itself. Nothing leaves the phone, and it works in airplane mode. - Gemma 4 E2B, Gemma 4 E4B, Qwen3.5 2B and Qwen3.5 4B (all Apache-2.0), each checked against its checksum - The engine is llama.cpp's official Android build, fetched with the first model (about 33 MB once installed) - Runs on the NPU (Snapdragon 8 Gen 2 and newer), the GPU or the CPU; Auto takes the fastest that starts …View on GitHub ↗
FAQ
Does Buddy itself cost anything?
What permissions does setup require?
What data leaves the phone?
Can I use it on an iPhone?
Do I need a computer or root access?
Is Buddy for Android open source?
Is Buddy for Android still maintained?
Who makes Buddy for Android?
What is Buddy for Android's official website?
Thanks — we use this to improve the page.