Install Fruitctl for Junie and IntelliJ

Fruitctl's junie adapter installs MCP configuration and the shared skill for Junie CLI or the standalone Junie IDE plugin. Junie inside IntelliJ's AI Chat has a separate IDE configuration step. These are experimental frontends; installation and a successful doctor check do not establish desktop acceptance.

Choose the frontend and prepare the connection

The first two destinations follow Junie's MCP documentation. The AI Chat steps follow Junie in AI Assistant. In an ACP client, /mcp lists configured servers and their source/status; it does not open the CLI installation assistant or edit configurations.

The standalone menu is build-dependent. Although the standalone guide still names Settings → Tools → Junie → MCP Settings, the Marketplace release metadata for 261.2144.120 and 262.2144.120, reviewed on 2026-10-07, includes its removal in the 2xx.1966.xx release notes. Use the documented JSON files and the MCP listing/status actually available in the selected frontend; do not assume the legacy menu exists. The AI Chat configuration path above is the separately documented AI Assistant surface.

Before installing, have these prerequisites ready:

• An existing Junie installation and its normal account or model-provider authentication. In IntelliJ, install/activate AI Assistant and select Junie in AI Chat, or install the separate Junie plugin for its standalone window.
• A selected model/provider that accepts MCP image content for the observation/action/observation journey. Tool registration and a text-only invocation do not establish vision capability.
• A named Fruitctl target profile on the Darwin controller, a private controller-local credential file, and owner-enabled target Screen Sharing. The complete configuration examples are in Home Manager.
• A running, configured Darwin broker. A Linux seat also needs its configured SSH relay and socket. The MCP subprocess runs on the machine executing Junie; that machine must have the runtime and the intended broker/relay connection.
• A writable project configuration, or the owning Home Manager configuration surface for store-managed files. Existing settings must remain intact.

The bootstrap installs the runtime and configuration. It does not install Junie or IntelliJ, create the target profile or credentials, start the broker/relay, enable Screen Sharing, or grant macOS consent. The public preview excludes the optional Host application and purple indicator.

One project-scope terminal paste

Run this in the root of the project you intend to open in Junie. Replace desktop with your already-configured target profile. This uses the immutable alpha.4 runtime preview and checks the pinned bootstrap script before execution. The first invocation previews the operation; the second installs it.

(
  FRUITCTL_TARGET_PROFILE='desktop'
  FRUITCTL_SOURCE_SHA='e0f4d064b076e58b33856d877e0e1c9266f55177'
  FRUITCTL_RELEASE_TAG='v0.1.0-alpha.4'
  FRUITCTL_SCRIPT=$(mktemp)
  trap 'rm -f "$FRUITCTL_SCRIPT"' EXIT
  curl --disable --fail --silent --show-error --location --proto '=https' --proto-redir '=https' \
    "https://raw.githubusercontent.com/xoxd-ai/fruitctl/$FRUITCTL_SOURCE_SHA/scripts/install.sh" \
    -o "$FRUITCTL_SCRIPT" &&
  FRUITCTL_ACTUAL_SHA=$(if command -v sha256sum >/dev/null 2>&1; then
    sha256sum "$FRUITCTL_SCRIPT"
  else
    shasum -a 256 "$FRUITCTL_SCRIPT"
  fi) &&
  [ "${FRUITCTL_ACTUAL_SHA%% *}" = '6e05a7a62f57491cfbda1f8d9212291f6237f04f4ac396bc04d183b88747cd58' ] &&
  sh "$FRUITCTL_SCRIPT" --agent junie --scope project \
    --version "$FRUITCTL_RELEASE_TAG" --target "$FRUITCTL_TARGET_PROFILE" --dry-run &&
  sh "$FRUITCTL_SCRIPT" --agent junie --scope project \
    --version "$FRUITCTL_RELEASE_TAG" --target "$FRUITCTL_TARGET_PROFILE"
)

The bootstrap needs curl, tar, awk, mktemp, and sha256sum or shasum; Node.js is bundled. For user scope, replace both --scope project occurrences with --scope user. This bootstrap selects project scope from the current directory; it has no --project-dir option. The installed CLI supports that option when subsequently selecting another project.

Use the receipt's recorded project path for subsequent lifecycle commands. The bootstrap uses the physical current directory. Alpha.4's installed CLI refuses an alternate spelling through a project symlink and names the recorded path, including during dry-run. Historical alpha.3 could create another receipt for that alias; alpha.4 does not migrate those older records automatically.

Inspect the JSON result. installed records the versioned runtime, MCP entry and skill link. declarative-required means configuration was retained: apply its fragment through Home Manager or the owning configuration surface. For an installed result, run from the same project:

"$HOME/.local/bin/fruitctl" doctor --agent junie --scope project

The generated mcpServers.fruitctl entry uses the absolute versioned executable with args: ["mcp", "--target", "desktop"] for the selected profile. An optional installed-CLI --config value becomes a FRUITCTL_CONFIG_PATH pathname; no profile contents or VNC password are embedded in MCP configuration. Junie's mutable enabled/disabled state is preserved on an owned entry update.

For AI Chat, copy the returned ideSettingsSnippet into the IDE's MCP settings using the documented STDIO JSON surface, then enable Pass custom MCP servers. The installer does not edit internal JetBrains XML. Adding this IDE entry is separate from the project JSON install.

Observed alpha.4 installation scope

The anonymous Linux x64 stock Junie project bootstrap passed, followed by installed help, doctor and reinstall. A continuation initialized the MCP SDK against Junie's installed MCP launcher and called listTools, discovering all four tool schemas without executing a tool action. An alternate-project-alias dry-run was refused, and uninstall restored original unrelated configuration bytes and modes. The 3,671 Linux runtime payload files were preserved, with the generated identity marker accounted for separately. Reinstall retains append-only installer history and its directory; the preservation claim covers configuration and runtime payloads.

On Darwin ARM64, alpha.4 separately passed an offline Junie project lifecycle through the bundled installer API and an installed SDK/shared-broker journey through the unchanged signed controller to a loopback RFB fixture. Four complete 96×53 frames, one reconnect and releases through task_complete and task_failed passed as synthetic protocol evidence. The Darwin stock public bootstrap remains pending.

Image rendering, desktop input, release and reconnect in an actual Junie conversation remain pending, as does separate IntelliJ acceptance. Historical alpha.2 seven-adapter and alpha.3 Claude/API checks do not qualify the alpha.4 Junie frontend.

Versions, skills and project instructions

Official documentation was reviewed on 2026-10-07. The standalone plugin guide lists IntelliJ Ultimate 2024.3.2 and Community 2025.1 as installation floors; the Marketplace changelog also records deprecation of the 2024.3/2025.1 lines. Those floors therefore do not establish compatibility with the latest plugin. Record the exact IDE and plugin build selected by Marketplace. Optional Junie CLI integration with a running IntelliJ instance requires IntelliJ 2026.1 or newer, the Junie plugin, and the same project path. That connection is separate from configuring Fruitctl MCP.

The installer links .junie/skills/fruitctl under the chosen scope. Current Junie skills documentation describes CLI and IDE support; the AI Assistant 2026.2 agent matrix, dated 2026-08-05, lists a narrower Junie skill surface. Verify discovery in the exact frontend rather than assuming that one document qualifies every build. For CLI, use /skills to check that fruitctl loads and is enabled, then /fruitctl or $fruitctl to request it. The documented folder locations do not by themselves prove that a particular frontend follows the installed symlink.

Keep the project's existing AGENTS.md. The standalone plugin's .junie/AGENTS.md, if already present, takes precedence over root instructions; Fruitctl does not create or replace it. Existing custom guideline settings also remain under the project owner's control.

Verify the real frontend journey

Use the existing Junie session and normal external-tool approval settings. Configuration registration is only the first stage:

1. Check MCP connection status in the selected frontend. In CLI, /mcp should show Fruitctl active with the expected project/user source. Confirm the actual discovered tool schemas, including vnc_command, action_queue, task_complete, and task_failed.
2. Request a fresh, complete desktop observation through Fruitctl. Inspect the image actually delivered in that conversation; a JSON image descriptor or a doctor result does not prove that Junie received useful pixels.
3. With the target owner's authorized test scene, make one reversible action, capture its visible result, release the session, and verify reconnect. Follow the installed skill and discovered schemas. Stop on an uncertain result rather than replaying input.
4. Record the exact release, Junie version, IDE/plugin build if applicable, execution host, capture mode and observed results. CLI, standalone plugin and AI Chat require separate records. Keep credentials and private screen data out of public evidence.

If an IDE cannot discover the skill, supply its reviewed instructions in that session and record skill discovery as incomplete. A passed configuration check or MCP listing does not change the frontend's experimental status; qualification requires the observation/action/observation and session-lifecycle results.