Firebase MCP Server

The Firebase MCP Server connects your AI assistant to your Firebase projects and to your app's code. MCP, the Model Context Protocol, is an open standard that lets assistants such as Claude Code, Cursor or GitHub Copilot in VS Code use external tools. Firebase is Google's app development platform, with services like the Cloud Firestore database, Firebase Authentication for sign-in, Crashlytics for crash reports and Hosting. The server comes straight from the Firebase team and ships inside the official Firebase CLI, the firebase-tools command line tool. You start it locally with npx -y firebase-tools@latest mcp. As long as Node.js and npm are available, there is nothing else to install.

What is the Firebase MCP Server good for?

Building an app on Firebase usually means bouncing between your editor, the Firebase console and a terminal. The MCP server brings that work into your assistant's chat. You can create a project, initialize services, query documents in Firestore or disable users in Authentication without opening the console. Debugging is where it really pays off: the assistant reads open Crashlytics issues and their events, checks Cloud Functions or App Hosting logs, and then suggests a fix right in your code. Because the server also knows the project directory with its firebase.json, it sees configuration and source code side by side.

Feature groups at a glance

According to the documentation, the tools are organized into feature groups. The core group is always on and covers sign-in, project and app details, SDK configuration, security rule validation, initialization and deployment. On top of that come auth for user management and custom claims, firestore for documents, queries, indexes, databases and backups, storage for download URLs, realtimedatabase for reads and writes, dataconnect for schemas and GraphQL operations, crashlytics, apphosting, functions, messaging for sending push notifications through Firebase Cloud Messaging, and remoteconfig for publishing or rolling back templates. The --only flag limits the tools to the groups you actually need, which keeps the tool list short and lowers the risk. The --dir flag pins the project directory.

Prompts, resources and documentation

Besides tools, the server ships ready-made prompts. firebase:init sets up services, including backend and AI features, and firebase:deploy kicks off a deployment. For Firestore and Cloud Storage, there are prompts that generate security rules along with unit tests. Resources hand the assistant guidance, for example on investigating Crashlytics issues or setting up Firestore. Developer Knowledge tools also let it search Google's official developer documentation, so it works from current information rather than stale training data.

Sign-in, security and data flow

The server runs on your machine and uses the same credentials as the Firebase CLI: your signed-in account or Application Default Credentials. That gives the assistant the same rights you have, including deletes for Firestore documents, databases, indexes and backups. Review every tool call before you approve it, and start with a test project. According to Google, some Firestore tools pass requests on to remote services. A few tools also use Gemini in Firebase. The README warns that its output can sound plausible and still be wrong, and that personal data should not go into it.

Cost and license

The server is part of the Firebase CLI and, like the CLI, is published on GitHub under the MIT license. There is no fee for the server itself. Regular Firebase plans apply: the free Spark plan with no payment method required, and the Blaze plan with pay-as-you-go billing for more services and higher quotas. Whatever the assistant creates or reads in your project is billed like any other usage.

Who is the Firebase MCP Server for?

The server suits app developers who already use Firebase and work with AI in their editor. It saves trips to the console, connects Crashlytics crash data with your code and takes over routine jobs like rules, indexes and deployments. Compared with the MCP servers from Supabase or Appwrite, it covers a particularly wide range, since it handles crash reports, push notifications and Remote Config alongside the database and auth. Claude Code and Cursor also get a plugin with agent skills, and the Gemini CLI has its own extension. If you do not use Firebase, it has nothing to offer you.

Node.js and npm, a Google account with access to a Firebase project (signed in through the Firebase CLI or Application Default Credentials) and an MCP client with stdio support, such as Claude Code, Claude Desktop, Cursor, VS Code with Copilot, Windsurf, Cline or Antigravity.

The Firebase MCP Server is part of the Firebase CLI and gives your AI assistant access to Firestore, Authentication, Crashlytics, Cloud Functions, Remote Config and other Firebase services. It runs locally over stdio. This guide covers setup in common clients and how to keep access sensibly limited.

Step 1: Prerequisites

You need Node.js and npm. The server uses the same credentials as the Firebase CLI, so sign in beforehand with npx -y firebase-tools@latest login, or provide Application Default Credentials. Alternatively, the assistant can sign you in later through the server's login tool.

Claude Code

The official plugin is the recommended route, because it adds agent skills on top of the server. Run claude plugin marketplace add firebase/firebase-tools, then claude plugin install firebase@firebase. If you only want the server, use claude mcp add firebase npx -- -y firebase-tools@latest mcp. Run claude mcp list to check that firebase shows up as connected.

Cursor, Claude Desktop and Windsurf

These clients share the same JSON entry. Under mcpServers, add a server called firebase with command npx and args -y, firebase-tools@latest and mcp. In Cursor, the entry goes into .cursor/mcp.json for one project or ~/.cursor/mcp.json for all of them. You can also install the Firebase plugin from the Cursor Marketplace instead. In Claude Desktop, open the config via Settings, Developer and Edit Config. Windsurf reads ~/.codeium/windsurf/mcp_config.json.

VS Code with Copilot

Create the file .vscode/mcp.json in your project. Under servers, add an entry called firebase with type stdio, command npx and the same args. To make the server available in every project, add it to your user settings under mcp and servers instead.

Gemini CLI and Firebase Studio

The Gemini CLI has its own extension, which you install with gemini extensions install https://github.com/firebase/agent-skills. In Firebase Studio, add the server to .idx/mcp.json.

Limit the tools

Append the --only flag to the args with a comma-separated list, such as auth,firestore,storage. The assistant then only sees the core group plus those areas. Use --dir with an absolute path to set which directory with a firebase.json counts as the project. That helps when your client does not start inside the project folder.

Work safely

The assistant acts with your full rights, including deleting Firestore data, sending push notifications and running deployments. Approve tool calls one at a time instead of allowing them across the board. Try new routines in a separate test project first. Tools that rely on Gemini in Firebase can give wrong answers, so check their output before you use it.

claude mcp add firebase npx -- -y firebase-tools@latest mcp

The same credentials as the Firebase CLI: an account signed in with firebase login, or Application Default Credentials. The core group also includes tools for signing in and out.

Everything the signed-in account is allowed to do: create projects and apps, initialize and deploy services, read, change and delete Firestore documents, databases, indexes and backups, manage users, send push notifications, publish Remote Config templates and read logs and Crashlytics data.

The server runs locally and talks directly to Firebase and Google Cloud APIs. Some Firestore tools pass requests on to remote services, and a few tools use Gemini in Firebase. Data that shows up in the chat also reaches the respective AI provider.

High. The assistant acts with your full rights and can delete production data, disable users, send push notifications to real users or trigger deployments. A crafted text with hidden instructions can set off such actions. Approve tool calls one by one, narrow them with --only and work in a test project first.