AndroidHarness is a coding agent that lives on your phone. It works on real projects straight from the device: reads and edits files, runs shell commands, drives git, controls a browser, and runs MCP tools. No PC required.

Open source under the MIT license: https://github.com/Sanuu7/AndroidHarness

THE AGENT RUNTIME

Chat with a coding agent: streaming responses, thinking blocks, and interactive tool cards. Attach skills, code, and PDFs. Queue follow-ups during runs or approve actions from notifications. Checkpoints survive app restarts. Dual planning inspects and plans before execution. Foreground service keeps runs alive with screen off.

57 BUILT-IN TOOLS

- Files: read_file, write_file, edit_file, multi_edit (fuzzy find-replace), apply_patch with rollback, grep, search_files, file_info, batch copy, move, delete.
- Shell: shell with timeouts and process group cleanup, background daemon manager (shell_background, bg_list, bg_kill).
- Git: git_status, git_diff, git_log, git_show (patch view), git_commit, git_branch, git_branch_manage, git_checkout, git_push with upstream repair, git_pull.
- System & Diagnostics: read_logcat with package and tag filters, doctor self-test suite, env_status probe.
- Vision & Web: read_image for vision models, PDF text extraction, web_search (keyless, Brave, Tavily), web_fetch, http_request with binary detection.
- Agent & Memory: memory_read, memory_write, memory_search, todo_write live tracking, ask_user prompts, task (parallel subagents), pkg_search, pkg_install, pkg_list.

MODEL CONTEXT PROTOCOL (MCP)

Connect external tool ecosystems through MCP:
- Connect servers over stdio, HTTP, and Server-Sent Events (SSE).
- Import configs by pasting Claude Desktop or MCP JSON.
- Servers expose custom tools dynamically to the agent runtime.
- Built-in OAuth 2.0 via Chrome Custom Tabs for authenticated services.
- Per-workspace approvals and HMAC integrity validation protect execution.

AGENT BROWSER (13 TOOLS)

Headless browser automation across 13 tools: browser_navigate, browser_click, browser_type, browser_scroll, browser_wait_for, browser_eval, browser_get_logs, browser_screenshot, browser_back, browser_forward, browser_refresh, browser_get_url, browser_get_dom. Preview web apps with DevTools, device frames, and one-tap bug repair back to the agent.

FILES, EDITOR, TERMINAL

Workspace file manager with multi-select and batch operations. Code editor with syntax highlighting (Kotlin, Java, Python, JS, TS, HTML, Shell), line numbers, unlimited undo, regex search, and encoding-preserving saves. Side-by-side visual diff viewer. Interactive terminal sessions.

AUTOMATIONS AND BUILD

Scheduled background automations via WorkManager: schedule planner, preset intervals, run history, alerts. Build & Test dashboard for Gradle commands and test suites with live logs and one-tap agent fixes for failed builds.

MODELS AND CACHING

Keyless access via built-in Harness provider, or bring API keys for Anthropic, Gemini, OpenAI, or custom URLs. Live catalog with latency probes, exact pricing, and real-time cost tracking. Prompt cache monitoring: hits, misses, read/write tokens, hit rates. Thinking ladders from off to ultra.

SAFE BY DESIGN

Execution tiers: Shizuku, Linux toolchain (bash, git, python, node), toybox sh fallback. Shell policy confines operations to workspace. Secret redactor scrubs API keys. Four permission modes: confirm all, confirm risky, full auto, full access. Biometric app lock.

SKILLS AND MEMORY

23 bundled markdown skills (planning, TDD, debugging, architecture) plus custom skills. Live todo tracking, repo map codebase indexing, Groq Whisper voice input, session forking, persistent memory.

GETTING STARTED

1. Grant storage access for project files.
2. Use built-in keyless provider or enter an API key in Settings.
3. Optional: connect Shizuku for privileged shell execution.

Zero analytics, ads, or trackers. Data flows only to configured endpoints. Report issues on GitHub.