Local-first Android development MCP server that provides tooling support for Gradle, adb, logcat, lint, crash triage, Perfetto, and visuals. It is described as focused on Android workflows, combining build, device interaction, debugging output, analysis, and performance/diagnostics capabilities in one context-aware interface.
🛠️ Key Features
Gradle integration
adb support
logcat access
lint support
Crash triage capabilities
Perfetto support
Visual tooling
🚀 Use Cases
Running and managing Android builds via Gradle
Interacting with connected Android devices using adb
Reading and analyzing runtime logs with logcat
Performing code checks with lint
Investigating crashes and triaging crash information
Capturing and analyzing performance traces with Perfetto
Using visuals related to Android development workflows
⚡ Developer Benefits
Centralized access to multiple Android development tools
Context focused on Android debugging, analysis, and performance data
⚠️ Limitations
Available tool list details are limited to the categories named in the description.
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.
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 (15 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, and build diagnose.
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
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 andconfirmDestructive=true on the call.
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.
Prerequisites: JDK 17+ on PATH, Node 18+ (install-time only for the launcher). Optional:
Android adb / platform-tools for device tools; set ANDROID_HOME or configure adbPath in the
user policy (~/.droidagentkit/policy.yaml).
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. The launcher always runs serve-mcp — for init /
install-mcp / audit use java -jar on the downloaded jar or build from source.
bash
npx -y @droidagentkit/launcher --version
Point your agent's MCP config at it directly, e.g. for Claude Code:
bash
claude mcp add droidagentkit -- npx -y @droidagentkit/launcher
Building from source
If you're contributing, or want to run a version that isn't released yet:
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:
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."
"What's in the app's SQLite schema?" (with storage + app_data_read enabled)
"Capture a Perfetto trace and summarize frame jank." (with perfetto enabled)
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, then tools/list).
The agent calls android_* tools via tools/call. You don't call them by hand.
--project auto resolves the active project from CLAUDE_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.
Run droidagent --help or droidagent <command> --help for flags.
Generating a config file
bash
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.
