maestro-mcp-setup

name	maestro-mcp-setup
description	Enable the Maestro CLI's built-in MCP server (`maestro mcp`, STDIO) for this machine, robustly. Detects the Maestro CLI and a usable Java runtime, installs/guides what is missing, and registers the MCP server at LOCAL (per-machine, uncommitted) scope with an absolute command path and injected JAVA_HOME/PATH so the spawn never dies on a non-login PATH. Use when an Expo project needs Maestro device-automation tools (list_devices, inspect_screen, take_screenshot, run flows). Never registers at project/committed scope — that reintroduces the fleet-wide "-32000" failure.
allowed-tools	["Bash","Read"]

Maestro MCP Setup

Wire the Maestro CLI's built-in MCP server (maestro mcp, STDIO) into the current coding agent for this machine only, in a way that actually works.

Why this skill exists (read before changing it)

The Maestro MCP server is intentionally NOT in the distributed plugin .mcp.json. A static, committed entry always attempts to spawn, and coding agents spawn stdio servers with a non-login PATH. So an always-on entry fails visibly — Claude Code shows Failed to reconnect … -32000 every session — on any machine missing the Maestro CLI or a resolvable Java runtime, which is most of a typical fleet (CI runners, unattended QA/intake agent boxes). That exact regression shipped once (expo plugin commit b12d5d3) and was reverted (b1f3efd).

This skill fixes both failure modes without poisoning the fleet:

1. Missing CLI / Java → detect, and install or guide install.
2. Installed but PATH-less spawn → register with an absolute command path plus injected JAVA_HOME / PATH env, so the server finds Java regardless of the spawn's inherited PATH.

Registration is always --scope local (per-machine, uncommitted). Never use --scope project / a committed .mcp.json entry — that is the always-on form that reds out every other machine.

Steps

1. Locate the Maestro CLI

MAESTRO_BIN="$(command -v maestro || true)"
[ -z "$MAESTRO_BIN" ] && [ -x "$HOME/.maestro/bin/maestro" ] && MAESTRO_BIN="$HOME/.maestro/bin/maestro"
echo "maestro: ${MAESTRO_BIN:-NOT FOUND}"

If not found, install it (official installer — confirm with the user first, since it writes to ~/.maestro):

curl -fsSL "https://get.maestro.mobile.dev" | bash
MAESTRO_BIN="$HOME/.maestro/bin/maestro"

Resolve to an absolute path — a bare maestro in the registration relies on the very PATH that is missing at spawn time.

2. Locate a usable Java runtime

Maestro needs a JDK (8+). Resolve JAVA_HOME honoring the project's version manager — do not install a parallel global JDK when one is already managed:

# Prefer an already-active/managed Java before installing anything. Covers the
# common version managers (mise, asdf, SDKMAN) plus an ambient install.
if command -v mise >/dev/null 2>&1 && mise which java >/dev/null 2>&1; then
  JAVA_BIN="$(mise which java)"
elif command -v asdf >/dev/null 2>&1 && asdf which java >/dev/null 2>&1; then
  JAVA_BIN="$(asdf which java)"
elif [ -x "${SDKMAN_DIR:-$HOME/.sdkman}/candidates/java/current/bin/java" ]; then
  # SDKMAN's `sdk` is a shell function absent in non-login shells; use its stable
  # `candidates/java/current` symlink directly instead.
  JAVA_BIN="${SDKMAN_DIR:-$HOME/.sdkman}/candidates/java/current/bin/java"
elif command -v java >/dev/null 2>&1; then
  JAVA_BIN="$(command -v java)"
fi
echo "java: ${JAVA_BIN:-NOT FOUND}"

# Derive JAVA_HOME by asking the JVM itself — NOT `dirname "$JAVA_BIN"/..`, which
# is wrong when `java` is a mise/asdf shim or a symlink whose parent is not a JDK
# root. Fall back to the bin/.. parent only if the JVM query yields nothing usable.
if [ -n "$JAVA_BIN" ]; then
  JAVA_HOME="$("$JAVA_BIN" -XshowSettings:properties -version 2>&1 \
    | sed -n 's/^[[:space:]]*java\.home = //p' | head -n 1)"
  { [ -z  ] || [ ! -x  ]; } &&
    JAVA_HOME=

If no Java is found: install through the project's version manager when present (mise use -g java@21, asdf install java …, or sdk install java), or guide the user to install a JDK. Prefer the managed path; only fall back to a global install with explicit consent. Re-resolve JAVA_HOME afterward.

3. Register at local scope with injected env

Build a PATH that puts the Maestro and Java bin dirs first, then register for the current coding agent. $JAVA_BIN/$MAESTRO_BIN are absolute.

Claude Code:

# Refuse to register a broken server — both toolchain paths must resolve and be
# executable, or `claude mcp add` would create the exact -32000 entry this skill
# exists to avoid.
[ -n "$MAESTRO_BIN" ] && [ -x "$MAESTRO_BIN" ] || { echo "Maestro CLI unresolved — abort"; exit 1; }
[ -n "$JAVA_HOME" ] && [ -x "$JAVA_HOME/bin/java" ] || { echo "Java runtime unresolved — abort"; exit 1; }

# Use the real JDK bin (from JAVA_HOME), not the shim dir, on PATH.
JBIN_DIR="$JAVA_HOME/bin"; MBIN_DIR="$(dirname "$MAESTRO_BIN")"
claude mcp add --scope local maestro \
  --env "JAVA_HOME=$JAVA_HOME" \
  --env "PATH=$JBIN_DIR:$MBIN_DIR:$PATH" \
  -- "$MAESTRO_BIN" mcp

Parity — same intent on other harnesses (local/per-machine scope, absolute command, injected JAVA_HOME/PATH); translate the registration verb only:

• Codex — codex mcp add maestro --env JAVA_HOME=… --env PATH=… -- "$MAESTRO_BIN" mcp (writes ~/.codex/config.toml; keep it out of the committed project config).
• Cursor — add the server to the user-scope ~/.cursor/mcp.json (not the project .cursor/mcp.json) with command, args: ["mcp"], and env: { JAVA_HOME, PATH }.
• OpenCode — add to the user config (~/.config/opencode/opencode.json) mcp block with the absolute command and environment.
• Antigravity (agy) / Copilot — add to the per-user MCP config with the absolute command + env; Copilot uses type: "local" for stdio servers.

In every case: local/user scope, never the committed project file.

4. Verify

claude mcp list        # or the harness equivalent

Confirm maestro shows connected. If it still fails, the usual cause is JAVA_HOME pointing at a JRE without bin/java, or a Maestro path that moved — re-resolve both and re-register. Do not "fix" it by dropping to a bare maestro mcp command; that reintroduces the PATH-less failure.

Guardrails

• Local scope only. If you catch yourself typing --scope project or editing a committed .mcp.json, stop — that is the reverted fleet-wide regression.
• Absolute paths + injected env are mandatory, not optional — they are the whole reason this is a skill and not a one-line doc.
• Installing the CLI or a JDK mutates the machine — confirm before installing, and prefer the project's version manager over a parallel global install.