Coding agents: tap-agent¶

Experimental

tap-agent, the server features it is built on (held connections, screen snapshots with refs, the event log) and the tap-events/1 export format are experimental: they may change in any release, including after 1.0, until this notice goes away.

tap-agent lets a coding agent (Claude Code, Cursor, …) drive a device through the same tap server your tests use. The agent reads the screen as text, acts on it, checks the result, and hands you a log of what it did. That covers checking a change on a real device, reproducing a bug, or exploring an app before writing a test for it.

It comes two ways, with the same steps and the same sessions: a CLI (one step per call, for agents that run shell commands) and an MCP server (the steps as tools). Both are built on the Python client. Every command is listed in the command reference.

Install¶

It needs the tap server (The tap server) and Python 3.10+. The Download bundle installs it with everything else; on its own, install the Python client first, then the agent tools, from the GitHub releases:

pip install https://github.com/NoamCohen48/tap/releases/download/client-python/v0.0.2/tap_e2e-0.0.2-py3-none-any.whl
pip install https://github.com/NoamCohen48/tap/releases/download/client-agent/v0.0.2/tap_agent-0.0.2-py3-none-any.whl
tap start

Connect your agent¶

Claude Code (MCP)Other MCP clientsCLI and a skill

claude mcp add tap -- tap-agent mcp

Any client that starts a stdio MCP server, such as Cursor or Claude Desktop:

{
  "mcpServers": {
    "tap": {"command": "tap-agent", "args": ["mcp"]}
  }
}

An agent that runs shell commands can use the CLI directly. tap-agent skill prints a SKILL.md that teaches it the flow below; save it where your agent reads skills or instructions (for Claude Code, .claude/skills/tap-android/SKILL.md; the skill is named tap-android).

mkdir -p .claude/skills/tap-android && tap-agent skill > .claude/skills/tap-android/SKILL.md

Then ask for something concrete: "On the emulator, open the Settings app, search for dark mode and tell me which settings come up."

Next¶

A session, step by step: attach, read the screen, act, check, and what to do when a step fails, from a real session.
Sessions to tests: export what the agent did and turn it into a test.
Commands: every command and MCP tool, with its options.