MCP server for React Native development — unified simulator/emulator control, Metro/Hermes runtime bridge, and structured native build diagnostics.
Strong foundation with 13 well-named tools following verb_noun convention (list_*, boot_*, launch_*, etc.). All tools have descriptions (avg ~120 chars, within baseline 34-392 range). Input schemas present and typed with Zod. Tool annotations (readOnlyHint, destructiveHint) correctly applied. Key gaps: output schemas not visible in source (only input schemas shown); parameter descriptions are present but some lack detail on constraints/formats; error handling patterns not evident in provided code; no pagination support for list_devices despite potentially large result sets.
Boot a simulator/emulator by id and wait until it is usable. No-op if already booted. For cold Android AVDs use the avd:<name> id from list_devices.
Execute a JavaScript expression inside the running React Native app and return the result. Use to inspect state (Redux/Zustand stores, globals) or trigger behavior. Promises are awaited via polling (Hermes limitation), up to timeoutMs. This can mutate app state — prefer read-only expressions when verifying. Requires Metro running.
Install a built app artifact on a booted device: a .app bundle (iOS simulator) or .apk (Android). Returns the app id for launch_app.
Launch an installed app on a booted device by bundle id / package name.
List iOS simulators and Android emulators/devices in a unified format. By default shows booted devices plus the newest-OS variant per device family; pass filter "all" for everything.
List debuggable React Native runtimes exposed by Metro. Call when a runtime tool reports TARGET_AMBIGUOUS or to check what is connectable. Requires Metro running.
Output schemas not documented in source code. Only input schemas visible; return types and structured output format not specified. LLMs cannot plan downstream tool calls without knowing what fields to expect.
list_devices lacks pagination despite potentially returning large result sets. No limit, offset, or cursor parameters.
Error handling patterns not visible in provided source. No evidence of recovery guidance, error classification (retryable vs fatal), or actionable error messages. Critical for agent resilience.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2025-06-18+ | v2 |
Open a URL on the device: custom schemes (myapp://...), universal links, or exp:// links. The primary way to drive app navigation from outside the app.
Read console logs from the running React Native app (buffered since the runtime bridge connected). Cursor-based: pass the previous nextCursor to read only new entries. Requires Metro running.
Trigger a full JS reload of the running React Native app (same as pressing 'r' in the Metro terminal). Requires Metro running.
Shut down a running simulator/emulator.
Capture a screenshot of a booted device. Returns the PNG path and (by default) the image itself for visual verification.
Stop a running app. Succeeds (terminated: false) if the app was not running.
Uninstall an app from a device. Destructive: removes the app and its data.
Parameter descriptions lack constraint details. E.g., timeoutSeconds has min/max (10-600) but description doesn't explain why or what happens if exceeded. Format constraints and validation rules should be in descriptions for LLM clarity.
evaluate_js tool accepts arbitrary JavaScript expressions. No evidence of sandboxing, input validation, or injection prevention. Agents could be tricked into executing malicious code.