DroidAgentKit
Give your AI coding agent real Android tools — Gradle, adb, logcat, lint, crash triage — through a local, permissioned MCP server instead of raw shell access.
DroidAgentKit helps you work on Android projects with AI agents. It runs on your machine, talks to your Gradle tree and (when you allow it) your emulator or device, and gives agents structured answers instead of raw shell output.
Android-first, deliberately. Cross-platform automation servers cover iOS too but stop at tapping the screen. DroidAgentKit goes the other way: deep into the Android build system and toolchain — Gradle task allowlists, lint, R8 crash triage, Perfetto traces, app storage, Compose visual regression. There is no iOS support and none planned.
Everything stays local. Commands go through allowlists. Output is redacted before it comes back to the agent. Artifacts land under build/droidagentkit by default.
What you get
Three pieces that work on their own or together:
MCP server — A Model Context Protocol server your agent can call. Inspect the project, run safe Gradle tasks, capture logs and screenshots, triage crashes, run lint, profile builds, and more. Optional groups add device diagnostics, bounded device control, Perfetto tracing, visual regression, read-only app storage inspection, and experimental emulator network capture.
Readiness auditor — Scores how agent-friendly your repo is, lists risks, and can generate AGENTS.md, a project skill file, and a starter .droidagentkit/config.yaml. Use --redact-public when you need a share-safe report.
Visual regression kit — Pixel-diff reports for Compose/UI captures, a Gradle plugin, and a JUnit rule. MCP tools can diff goldens, generate reports, and update goldens when you explicitly confirm.
MCP tools at a glance
The core group is on by default (18 tools): project inspect, allowlisted Gradle runs, device list, install/launch, logcat, screenshot, accessibility snapshot, report bundle, lint, crash triage, dependency check, build performance, test run, build diagnose, APK size analysis, UI element lookup, and android_doctor (an environment preflight you can run when something fails for no obvious reason).
Turn on more by exposing tool groups when you start the server (see security model):
| Group | What it adds |
|---|---|
device_read | Permission audit, dumpsys summaries, memory/battery, bugreport, streaming logcat jobs |
device_control | Emulator control, app uninstall/clear, intents, permissions, tap/swipe/type, tap-by-label, file push/pull, small action flows, flow recording |
perfetto | Trace capture and Trace Processor analysis (jank, CPU, contention, binder latency, …) |
visuals | Pixel diff, visual report, golden updates |
storage | Read-only SQLite, SharedPreferences, and file tree for debuggable apps |
network_experimental | Emulator-only mitmproxy capture + redacted HAR query (requires your own debug CA) |
Dangerous actions (uninstall, clear data, proxy install, golden overwrite, …) need the right capability in config and confirmDestructive=true on the call. The capability is the real boundary — confirmDestructive guards against an accidental call, not a hostile one, since the agent supplies it. See the threat model.
Report bundles include a capability summary: which groups are exposed, which capabilities are enabled, and what you still need installed (adb, trace processor, mitmproxy, …). That summary is informational — it does not change your readiness score.
Full tool list and capability IDs: docs/security-and-permissions.md.
Quick start
Prerequisites: Node 18+ (install-time only for the launcher). A JDK 17+ is used if you have
one — on PATH, in JAVA_HOME, or named by DROIDAGENT_JAVA — and otherwise a pinned Eclipse
Temurin JRE is downloaded once, SHA-256 verified, and cached under ~/.droidagentkit/jre.
Optional: Android adb / platform-tools for device tools; set ANDROID_HOME or configure
adbPath in the user policy (~/.droidagentkit/policy.yaml). Run doctor to check all of this.
The @droidagentkit/launcher npm package downloads and caches the matching droidagent-cli jar from
GitHub Releases the first time it runs,
verifying its SHA-256 before executing it.
npx -y @droidagentkit/launcher --version
Run it bare to start the MCP server, or pass any CLI command straight through:
npx -y @droidagentkit/launcher doctor
npx -y @droidagentkit/launcher init --profile device-control
npx -y @droidagentkit/launcher audit --project /path/to/android --write-agents
Point your agent's MCP config at it directly, e.g. for Claude Code:
claude mcp add droidagentkit -- npx -y @droidagentkit/launcher
Or install with one click:
Other install paths
brew install iVamsi/android-agent-kit/droidagent
Every release also attaches an .mcpb bundle. Drop it into an MCP host that supports one-click
install and the server registers itself. The bundle carries the launcher only — the JVM server, and
a verified Temurin JRE if you have no JDK, are fetched on first run — so the download is ~10 KB
rather than ~160 MB.
Building from source
If you're contributing, or want to run a version that isn't released yet:
./gradlew :cli:installDist
./cli/build/install/droidagent/bin/droidagent install-mcp
Register it once for your agent (Codex, Claude Code, Cursor, Zed, VS Code, Android Studio, …).
Preview without writing files with install-mcp --dry-run. For Android Studio across many
projects under one folder:
droidagent install-mcp --targets android-studio --projects-root ~/Developer/StudioProjects
Details and per-host config paths: docs/easy-mcp-installation.md.
Security model + full tool tables: docs/security-and-permissions.md.
CLI reference + troubleshooting: docs/cli-reference.md, docs/troubleshooting.md.
Try it in chat
After install-mcp, open an Android project and ask things like:
- "Inspect this Android project with DroidAgentKit."
- "Run the safe unit tests for the app module."
- "Audit this repo for agent readiness and write AGENTS.md."
- "Capture logcat and triage the crash." (release builds are de-obfuscated automatically when an R8 mapping is present)
- "Why did the APK grow since the last build?"
- "What's in the app's SQLite schema?" (with
storage+app_data_readenabled) - "Capture a Perfetto trace and summarize frame jank." (with
perfettoenabled) - "Record what you just did as a flow I can replay." (with
device_controlenabled)
You describe the goal; the agent picks tools from the list it was given at startup.
How it works
- Your agent starts
droidagent serve-mcp --transport stdio --project auto(or talks to the localhost HTTP endpoint Android Studio uses on macOS). - MCP handshake (
initialize, thentools/list). - The agent calls
android_*tools viatools/call. You don't call them by hand. --project autoresolves the active project fromCLAUDE_PROJECT_DIR,CODEX_WORKSPACE/CODEX_PROJECT_DIR,GEMINI_PROJECT_DIR/GEMINI_WORKSPACE,CURSOR_PROJECT_DIR/CURSOR_WORKSPACE,PWD, then cwd — one user-wide install, many projects.
Optional resources and workflow prompts are available on non–Android Studio hosts. Tool listings include hints like readOnlyHint and destructiveHint where relevant.
CLI without an agent
droidagent inspect --project /path/to/android --format markdown
droidagent audit --project /path/to/android --write-agents
droidagent audit --project /path/to/android --redact-public
droidagent devices --format json
droidagent serve-mcp --transport stdio --project auto
Run droidagent --help or droidagent <command> --help for flags.
Generating a config file
droidagent init # interactive: six yes/no prompts explaining risk per area
droidagent init --profile device-control # non-interactive, for scripted setup
droidagent init --profile full # everything, including storage and network capture
droidagent init --list-profiles # see all profile names without writing anything
Only the core tool group (safe, read-only) is enabled by default. droidagent init is the fastest way to
turn on more without hand-writing .droidagentkit/config.yaml — see
docs/security-and-permissions.md for what each group and capability
actually grants an agent.
Configuration
Put .droidagentkit/config.yaml in your Android project root to tune Gradle allowlists, adb permissions, capability flags, redaction patterns, and report output paths. The auditor can seed a safe default when you run audit --write-agents.
Status
Alpha. MCP tool names and schemas are treated as stable, but behavior will tighten as we learn from real use. No telemetry. No cloud dependency for core workflows.
More docs
- Security and permissions — capabilities, tool groups, destructive ops, worked prompt-injection scenarios
- Compatibility — what is stable, what may change, 1.0 exit criteria
- Easy MCP installation — per-host setup
- CLAUDE.md — build commands and architecture notes for contributors