Baepsae

io.github.oozoofrogPublished by GitHub account oozoofrogv6.3.0Updated Oct 24, 2026

Local MCP server for iOS Simulator and macOS app automation with a native Swift bridge

VerifiedSTDIODesktop onlyBrowser AutomationDeveloper Tools

Overview

AI-generated overview

A local MCP server that lets an assistant inspect and drive iOS Simulators and macOS apps through a native Swift bridge.

What you can do
Inspect and query the UI accessibility tree, then tap, type, swipe, scroll, drag and set values by identifier or label, including tab switching and dialog detection with a wait step for UI changes and a run_steps workflow tool for chaining actions. Send keyboard input, key sequences and key combos, touch input, and list or switch input sources. Manage simulators: list devices, take screenshots, record or stream video, open URLs, and install, launch, terminate or uninstall apps. Work with macOS apps and windows: list apps and windows, activate or focus a window, screenshot an app, right-click, use context menus and menu actions, read the clipboard, and watch notifications. Run diagnostics: a
When to use it
Useful when you are developing or testing iOS apps in the Simulator or automating macOS apps on your own Mac, and want an assistant to inspect the UI, drive taps and typing, capture screenshots or video, and manage app installs and launches. It is macOS-only, so it is not an option on Linux or Windows.
Requirements
macOS 14 or later with Xcode or Xcode Command Line Tools (for xcrun simctl and Swift 6+ to build the native binary), plus Node.js 18 or later; installable via npx or npm. Accessibility permission must be granted to the launching terminal, MCP client or runtime process (node), and the process may need a restart afterwards. A booted simulator is needed for simulator features.
Before you install
The server can control the mouse and keyboard of your Mac and interact with any app through the Accessibility API, so grant permission only to processes you trust. It can install, launch, terminate and uninstall apps on simulators, and typing on macOS targets temporarily overwrites the host clipboard before restoring it. It runs locally with no declared authentication or API keys.

Installation

In SourceWeft

  1. Open Baepsae in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

mcp-baepsae

[baepsae]

Baepsae (Vinous-throated Parrotbill) — A tiny Korean bird. Round, chubby, and constantly hopping around chirping. Known for its grit — even when a little bird tries to keep up with a stork, it never gives up. This project is small too, but it pecks away at your simulators tirelessly.

Local MCP server for iOS Simulator and macOS app automation with a TypeScript MCP layer and a Swift native bridge.

한국어 문서는 README-KR.md를 참고하세요.

Table of Contents

Prerequisites

  • macOS 14+
  • Xcode + iOS Simulator
  • Node.js 18+
  • Swift 6+

Platform Support

PlatformSupportedNotes
macOSYesPrimary platform. Required for iOS Simulator and Accessibility APIs.
LinuxNoNative binary depends on AppKit, CoreGraphics, and Accessibility frameworks.
WindowsNoNative binary depends on AppKit, CoreGraphics, and Accessibility frameworks.

Why macOS only?

The Swift native bridge (baepsae-native) uses macOS-specific frameworks (AppKit, CoreGraphics, Accessibility) to interact with iOS Simulator and macOS applications. These frameworks are not available on Linux or Windows. The TypeScript MCP layer also relies on xcrun simctl, which is part of Xcode Command Line Tools and only available on macOS.

Requirements summary:

  • macOS 14 or later -- required for iOS Simulator automation and Accessibility API access.
  • Xcode or Xcode Command Line Tools -- required for Swift 6+ compilation of the native binary and xcrun simctl commands.
  • Node.js >= 18.0.0 -- required to run the TypeScript MCP server.

Permissions

Accessibility permission is required for UI inspection and input automation features (use unified generic tools such as analyze_ui, tap, right_click).

The important detail is that permission usually needs to be granted to the automation host / runtime process, not to the target app you are automating.

Which process usually needs permission?

  • Direct native binary invocation
    • Example: baepsae-native ...
    • Most relevant entry: the baepsae-native binary itself, plus the terminal/shell app that launched it
  • Node / npx MCP runtime
    • Example: node dist/index.js, npx -y mcp-baepsae@latest
    • Most relevant entry: the runtime process (node), plus the terminal or MCP client app that launched it
  • Desktop / CLI MCP clients
    • Example: Claude Code, Codex CLI/Desktop, Gemini CLI
    • Relevant entries can include the MCP client app, the terminal host, and the runtime process depending on launch path

Recommended setup flow

  1. Open System Settings > Privacy & Security > Accessibility.
  2. Enable the terminal or MCP client app you actually use.
  3. Enable the runtime process if listed (node, bun, etc.).
  4. For direct native invocation, also enable the baepsae-native binary entry if it appears separately.
  5. If an entry is missing, click + and add it manually.

Important note

After granting permission, the launching process may need to be restarted before macOS applies the change.
If the error persists, quit and relaunch the terminal, MCP client, or runtime process that started mcp-baepsae.

For simulator targets, selector-based actions (tap / right_click with id or label) search in-app content by default. Set all: true to include Simulator chrome UI.

Install

Option A) npm (easiest)

bash
# Run directly without installingnpx mcp-baepsae@latest
# Or install globallynpm install -g mcp-baepsae

On macOS, the Swift native binary is built automatically during install. If Swift is not available, the server still works with simctl-based features.

If you plan to use UI inspection or input automation right away, grant Accessibility permission to the terminal / MCP client / runtime process you will actually launch the server with.

Option B) From source

bash
git clone https://github.com/oozoofrog/mcp-baepsae.gitcd mcp-baepsaenpm installnpm run build

MCP Setup (Recommended)

Use the installer script that calls each client's MCP install flow directly.

bash
bash scripts/install.sh --tool all

Client Matrix

ClientInstall PathScript TargetNotes
Claude Codeclaude mcp add--tool claude-codeSupports multi-scope via --claude-scope
Claude Desktopclaude mcp add --scope user--tool claude-desktopUser-level registration
Codex CLIcodex mcp add--tool codex-cliAlso works for Codex Desktop settings
Codex DesktopCodex CLI MCP config--tool codex-desktopUses same MCP registry as CLI
OpenCodeGlobal config (~/.config/opencode/opencode.json)--tool opencodeAuto-updated by installer
Geminigemini mcp add--tool geminiUser scope setup
Google AntigravityGemini-compatible MCP flow--tool antigravityUses Gemini MCP command path
GitHub Copilotcopilot or gh copilot session--tool copilotInteractive/session-based setup

For LLM

If you are an LLM agent setting up this MCP server, here is everything you need:

Before trying UI automation tools, remember that Accessibility permission usually belongs to the host/runtime process (node, terminal, MCP client), not the target app under automation.

Quick start (one command)

bash
# Register baepsae as an MCP server for all supported clientsbash scripts/install.sh --tool all

If you installed via npm instead of cloning the repo, use npx:

bash
# Claude Codeclaude mcp add baepsae -- npx -y mcp-baepsae@latest
# Codex CLIcodex mcp add baepsae -- npx -y mcp-baepsae@latest

Automation flags

bash
# Preview commands without executingbash scripts/install.sh --tool all --dry-run
# Verify environment and dependenciesbash scripts/install.sh --tool all --doctor
# Unregister from all clientsbash scripts/install.sh --tool all --uninstall

Runtime options

The installer supports multiple runtimes via --runtime:

FlagCommandWhen to use
--runtime node (default)node dist/index.jsLocal source build
--runtime npxnpx -y mcp-baepsae@latestnpm registry, no global install
--runtime bunxbunx mcp-baepsae@latestBun users
--runtime globalmcp-baepsaeAfter npm install -g mcp-baepsae

Manual Setup (Fallback)

Use this when you do not want to run scripts/install.sh.

Using npx (recommended for npm users)

bash
# Claude Codeclaude mcp add baepsae -- npx -y mcp-baepsae@latest
# Codex CLIcodex mcp add baepsae -- npx -y mcp-baepsae@latest
# Gemini CLIgemini mcp add --scope user --transport stdio baepsae npx -y mcp-baepsae@latest

When using npx, the relevant Accessibility entry is commonly the spawned node runtime plus the terminal / MCP client that launched it.

Using local build

bash
# Claude Code (project)claude mcp add --scope project --env="BAEPSAE_NATIVE_PATH=/ABS/PATH/native/.build/release/baepsae-native" baepsae -- node /ABS/PATH/dist/index.js
# Codex CLIcodex mcp add baepsae --env BAEPSAE_NATIVE_PATH=/ABS/PATH/native/.build/release/baepsae-native -- node /ABS/PATH/dist/index.js
# Gemini CLIgemini mcp add --scope user --transport stdio -e BAEPSAE_NATIVE_PATH=/ABS/PATH/native/.build/release/baepsae-native baepsae node /ABS/PATH/dist/index.js

When using a local build, check permission on both the runtime (node) and the app that launched it.
If you invoke baepsae-native directly for debugging, check permission on the native binary entry itself as well.

Project Structure

  • MCP server entry point: src/index.ts
  • Tool modules: src/tools/ (info, simulator, ui, input, media, system)
  • Shared utilities: src/utils.ts, src/types.ts
  • Native binary entry point: native/Sources/main.swift
  • Native command handlers: native/Sources/Commands/
  • Native binary output: native/.build/release/baepsae-native
  • TS tests: tests/mcp.contract.test.mjs, tests/unit.test.mjs, tests/mcp.real.test.mjs
  • Swift tests: native/Tests/BaepsaeNativeTests/

Commands

bash
npm run build       # Build TypeScript + native Swift binarynpm test            # Contract/integration testsnpm run test:real   # Real simulator smoke test (requires booted simulator)npm run test:real:preflight  # Environment diagnostics onlynpm run test:real:sim        # iOS simulator phases only (skips Phase 4)npm run test:real:mac        # macOS Safari phase onlynpm run verify      # test + test:realnpm run setup:mcp   # Alias for scripts/install.sh

MCP Tool Status

43 tools implemented end-to-end.

Official public MCP surface: unified generic tools

The public API surface is intentionally single-scheme: use unified generic tools with a target argument, rather than sim_* / mac_* names.

CategoryTools
UIanalyze_ui, query_ui, tap, tap_tab, type_text, swipe, scroll, drag_drop, wait_for_ui, detect_dialog, read_ui_value, set_ui_value, read_ui_param, hit_test, enumerate_ui
Inputkey, key_sequence, key_combo, touch, input_source, list_input_sources
Workflowrun_steps
Systemlist_windows, activate_app, screenshot_app, right_click, focus_window, context_menu_action, watch_notification
Simulator-onlylist_simulators, screenshot, record_video, stream_video, open_url, install_app, launch_app, terminate_app, uninstall_app, button, gesture
macOS/systemlist_apps, menu_action, get_focused_app, clipboard
Utilitybaepsae_help, baepsae_version, doctor

Target routing is explicit in the arguments: udid for simulator, bundleId / appName for macOS.

type_text policy

type_text accepts exactly one input source: text, stdinText, or file.

  • method: "auto" resolves to:
    • paste for simulator targets
    • keyboard for macOS targets
  • method: "paste" uses the simulator pasteboard for simulator targets and a temporary host clipboard replace/restore flow for macOS targets.
  • method: "keyboard" always types character-by-character.

When paste is used, simulator targets update the simulator pasteboard without touching the host clipboard, while macOS targets temporarily overwrite the host clipboard and restore it after submission. Successful responses report the input source, target kind, requested method, used method, paste transport, and any auto fallback that was applied.

tap_tab policy

tap_tab is semantic-first, geometry-last.

  • First it tries actionable descendants exposed under the tab bar.
  • If SwiftUI/Simulator does not expose real tab button descendants, it can use a semantic proxy row (for example, app-provided top navigation buttons that map cleanly to the same tabs).
  • Only if neither semantic path exists does it fall back to tab-bar geometry.

This matters because SwiftUI TabView in Simulator may expose the tab bar as a generic AXGroup text="Tab Bar" without actionable child buttons.

Usage Examples

Switch tabs by index with semantic-first fallback:

javascript
// Prefer this when your app exposes a real tab bar but individual tab buttons// are not consistently addressable via query_ui/tap selectors.tap_tab({ udid: "...", index: 1, tabCount: 3 })

Unified simulator app accessibility quickstart (inside app UI):

javascript
// 1) Launch your app in the target simulatorlaunch_app({ udid: "...", bundleId: "com.example.app" })
// 2) Inspect or search accessibility tree (in-app content scope by default)analyze_ui({ udid: "..." })query_ui({ udid: "...", query: "Login" })
// 3) Interact by accessibility identifier/labeltap({ udid: "...", id: "login-button" })
// Optional: include Simulator chrome/system UI in selector lookuptap({ udid: "...", label: "Home", all: true })

Open a URL (iOS Simulator):

javascript
// Open Naver Mobileopen_url({ udid: "...", url: "https://m.naver.com" })

Manage Apps (iOS Simulator):

javascript
// Install an appinstall_app({ udid: "...", path: "/path/to/App.app" })
// Launch Safarilaunch_app({ udid: "...", bundleId: "com.apple.mobilesafari" })
// Terminate Safariterminate_app({ udid: "...", bundleId: "com.apple.mobilesafari" })

macOS App Automation:

javascript
// List running macOS appslist_apps({})
// Take screenshot of a macOS appscreenshot_app({ bundleId: "com.apple.Safari" })

Troubleshooting

Accessibility permission checklist

  • The permission target is usually the automation host/runtime process, not the target app.

  • Run doctor first to inspect host process, parent process, native binary, booted simulator availability, and accessibility readiness in one place.

  • Check the error message for:

    • current host process
    • parent process
    • inferred launch mode
  • If you launched through npx / node, grant permission to the runtime and the launching terminal / MCP client.

  • If you launched baepsae-native directly, grant permission to the native binary entry and the launching terminal / shell app.

  • After changing permission, restart the launching process before retrying.

  • Invalid environment variable format on Claude setup:

    • Use current script (scripts/install.sh) or claude mcp add --env="KEY=value" ... format.
  • Missing native binary error:

    • Run npm run build and confirm native/.build/release/baepsae-native exists.
  • Accessibility permission error is ambiguous:

    • Current versions include host/parent process diagnostics and inferred launch mode in the error text so you can see which executable path likely needs permission.
  • Real smoke test diagnostics:

    • Run npm run test:real:preflight to print environment and capability diagnostics without executing the full suite.
    • Run npm run test:real:sim to focus on simulator-capability coverage, or npm run test:real:mac for the macOS Safari subset.
  • Multiple Simulator windows are open and selectors hit the wrong device:

    • Current versions scope simulator selectors to the target udid window first.
    • If you intentionally need Simulator chrome/system UI, use all: true.
  • OpenCode does not show baepsae:

    • Re-run bash scripts/install.sh --tool opencode --skip-install --skip-build and check ~/.config/opencode/opencode.json.
  • Copilot not auto-registered:

    • Copilot MCP flow is interactive/session-based. Re-run installer with --interactive.
  • Real smoke test skipped:

    • Boot an iOS simulator first, then run npm run test:real.

Source: README.md at commit 4e36492

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v6.3.0LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts