Garmin Simulator MCP
# Garmin Emulator MCP
```text
█████████
████████·······████████
█████·····················█████
LIGHT ▐ ████··┌─────────────────────┐··████ ▌ START
████····│ M C P │····████
████·····│ │·····████
███······│ > ready_ │······███
UP ▐ ████······│ │······████
███······│ │······███ ▌ BACK
████·····│ build run │·····████
████····│ │····████
DOWN ▐ ████··└─────────────────────┘··████
█████·····················█████
████████·······████████
█████████
```
Watch apps are awkward to test. The screen is tiny, there is no keyboard, and the simulator expects a human with a mouse. This puts the simulator behind an MCP server so an AI coding agent can run the same loop you would: build, look, press, look again, read the log.
## What it does
| Capability | What you get |
| :--- | :--- |
| Build | Compiles a Monkey C project with the Garmin SDK and loads it into the simulator. |
| Drive | Presses the physical buttons, taps, swipes and long-presses the watch screen. |
| Isolate | Runs the simulator on a hidden desktop, so it never takes over your screen. |
| Inspect | Screenshots of the watch display, the bezel, or an annotated debug view. |
| Read | App logs and crash output streamed back to the agent. |
| Check | SDK, device profiles, signing key and simulator state before every run. |
| State | Reads and seeds app storage and debug property overrides, resets data and moves the app clock. |
| Message | Sends a phone-app message into the app and reads the stored payload back. |
| Coverage | Builds and captures the same screen, and optional button or touch steps, across several devices. |
| Loop | Builds, runs and returns the first frame and log tail in one call. |
## The loop
| Step | What happens |
| :--- | :--- |
| 1 | The agent checks the setup and launches the simulator. |
| 2 | It builds your project and runs it on the selected device. |
| 3 | It presses through the screens you describe. |
| 4 | It looks at the result and reads the logs. |
| 5 | It repeats until the screen matches what you wanted. |
## Requirements
| Need | Why |
| :--- | :--- |
| Node.js 18 or later | Runs the server. |
| Java | Runs the Garmin compiler. |
| Connect IQ SDK Manager | Provides the SDK, simulator and device profiles. |
| A developer signing key | Signs your app for the simulator. |
| macOS 13 or Windows with PowerShell 5.1 | Supported hosts. Windows uses a hidden desktop by default; an unlocked desktop is only needed for the visible fallback. |
## Quick start
| Step | Command |
| :--- | :--- |
| Install | `npm install` |
| Build | `npm run build` |
| Connect | Point your MCP client at `dist/index.js`. |
Ask the agent to check the setup, launch the simulator, build your project and take a screenshot. Setup examples for common clients are in [AGENT-README.md](AGENT-README.md).
`npm install` can print install-script warnings from `sharp`, `esbuild` or `fsevents`. They are dependency build messages, not server errors; `npm run build` is the check that matters.
## Good to know
| Topic | Detail |
| :--- | :--- |
| One at a time | One simulator, one server. A second instance refuses to start so two agents cannot fight over the same device. |
| Focus | Button presses need an unlocked, interactive desktop. Screenshots, logs and menu actions keep working in the background. |
| Best effort | The watch screen is not exposed as native controls, so the server verifies each press by comparing screenshots. |
| Local state | Session logs and captures are written under your app data folder, not into the project. |
## Agent reference
[AGENT-README.md](AGENT-README.md) lists every MCP tool and its parameters.
## License
Apache-2.0.
TDQS
Scored across 46 tools
The broad domains are distinct (lifecycle, input, capture, debug, files), but several boundaries blur: debug_command overlaps state_read/state_seed/state_reset/set_time/phone_message, and reset_state/state_reset/debug_command reset all touch reset behavior. matrix_run vs. build+run vs. dev_loop, and capture_window/get_screenshot/diff_screenshot also require careful reading to choose correctly.
Names are consistently snake_case and mostly action-oriented, which makes the set readable. The pattern varies between verb_noun (get_screenshot, list_devices), noun_verb (session_start, state_read, simulator_status), and bare utility names (doctor, calibrate), but this is a minor deviation rather than a different convention.
At 46 tools, the surface is heavy for a simulator-control server, with many one-off diagnostic, matrix, window, and debug helpers. While the domain is genuinely broad, the count exceeds a well-scoped set and invites tool-selection overhead.
Coverage is strong across lifecycle, build/run, input simulation, screenshots/diffs, debug-channel state, file snapshots, windows, and diagnostics. Minor gaps remain around explicit baseline management (listing/deleting baselines) and app install/uninstall-style operations, but core workflows are well supported.