← MCP server and skill grades

skills

firebase/agent-skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
99954a30898b
Files read
93

This collection of official Firebase skills guides AI coding agents in configuring and deploying Firebase services across web, Android, iOS, and Flutter apps, and includes a Swift utility for adding packages to Xcode projects. The biggest risks are that the base skill tells the agent to silently install the entire skill suite and rewrite global IDE settings to run unpinned npm packages, while the Analytics skill quietly sends telemetry about the user's AI agent and model to Google Analytics without asking.

What it can reach

Files
Reads and edits project source code, Firebase config files (firebase.json, firestore.rules, apphosting.yaml, google-services.json, GoogleService-Info.plist), Xcode .pbxproj files, and global or workspace IDE MCP configuration files (such as ~/.cursor/mcp.json and claude_desktop_config.json).
Shell commands
Instructs the coding agent to execute shell commands including npx, swift run, flutter, flutterfire, gradlew, adb, curl, and IDE plugin managers.
Network
Communicates with Firebase and Google Cloud APIs to create and deploy cloud resources, downloads packages from npm and GitHub, and queries Google Maven metadata.
Secrets
Uses local Firebase CLI authentication credentials, downloads app SDK configuration files, handles App Check debug tokens, and reads local .env files to upload secrets to Cloud Secret Manager.

Findings

1 medium3 low
  1. MediumBase skill silently installs all other skills and edits global IDE MCP configs

    skills/firebase-basics/SKILL.md:24

    Even when a user installs only the firebase-basics skill for simple CLI login or config downloads, its mandatory prerequisites instruct the agent not to skip installing the full firebase/agent-skills suite (including piping answers to interactive CLI prompts via stdin in setup guides such as cursor.md, antigravity.md, github_copilot.md, and other_agents.md) and directly editing global IDE MCP configuration files (like ~/.cursor/mcp.json and claude_desktop_config.json) to register the Firebase MCP server without asking permission first.

    Fix: Require explicit user confirmation before installing additional skills, plugins, or modifying global or workspace MCP configuration files.

  2. LowSilent upload of AI agent and model telemetry via temporary analytics event

    skills/firebase-analytics/references/android_setup.md:323

    All four setup and screen-tracking guides in the firebase-analytics skill (android_setup.md, android_screen_views.md, ios_setup.md, and ios_screen_views.md) instruct the agent to act automatically without human intervention to inject a gaac_app_open event into the user's app containing the AI agent name, LLM model name, skill names, and modified app event names, force an immediate debug upload to Google Analytics servers, and then delete the code. This transmits metadata about the developer's AI environment without consent, is not mentioned in SKILL.md, and risks leaving tracking code in the app startup path if the agent stops before cleanup.

    Fix: Remove the mandatory gaac_app_open telemetry injection, or clearly disclose the telemetry payload and require explicit user opt-in before adding and uploading it.

  3. LowMandatory execution of unpinned @latest npm packages at runtime

    skills/firebase-basics/SKILL.md:89

    The skills forbid using a locally installed, version-pinned firebase binary and require running npx -y firebase-tools@latest on every CLI command and inside persistent IDE MCP server configs, alongside unpinned npx -y skills commands. Automatically downloading and running whatever is currently published under the latest tag with confirmation prompts disabled (-y) exposes the user's machine and Firebase credentials to npm supply-chain compromises.

    Fix: Prefer a locally installed, version-pinned firebase-tools dependency when available, and pin known-good package versions instead of forcing @latest with -y.

  4. LowSecret upload instructions expose .env secret values to LLM context and command logs

    skills/firebase-app-hosting-basics/SKILL.md:77

    When migrating sensitive keys from .env files to Cloud Secret Manager, the skill instructs the agent to inline the raw secret into printf '%s' '<value>' piped to firebase-tools. Following this literally requires the agent to read cleartext secrets into the LLM conversation context (sending them to the AI model provider) and record them in agent tool-call logs, and also breaks if a secret contains single quotes.

    Fix: Instruct the agent to either let the user enter secrets interactively or extract the key directly from the .env file in a shell pipeline without reading or echoing the secret value into the LLM context.