A Model Context Protocol (MCP) server for Firestore, enabling AI agents and LLMs to interact with Firestore databases securely through multiple transport protocols.
- Multiple Transport Protocols: Support for stdio, HTTP Streamable, and Server-Sent Events (SSE)
- Secure by Design: Uses Firebase Client SDK with Firestore Security Rules instead of Admin SDK
- Complete CRUD Operations: Get, set, add, delete, and query Firestore documents
- Type-Safe: Built with TypeScript and Zod schemas
- MCP Compatible: Works with any MCP-compatible client (Claude Desktop, etc.)
FireMCP uses the Firebase Client SDK instead of the Admin SDK for a critical security reason:
The Client SDK respects Firestore Security Rules, ensuring that AI agents can only access data they're explicitly permitted to. This prevents unauthorized access to sensitive resources, even if the AI behaves unexpectedly.
With the Admin SDK, an AI would have unrestricted access to all Firestore data, which could be a significant security risk in production environments.
- Bun v1.0 or higher
- A Firebase project with Firestore enabled
- Firebase Authentication configured with at least one user
git clone https://github.com/iamanishroy/firemcp.git
cd firemcpbun installCopy the example environment file and fill in your Firebase credentials:
cp .env.example .envEdit .env with your Firebase configuration:
# Required
FIREBASE_API_KEY=your-api-key-here
FIREBASE_PROJECT_ID=your-project-id
FIREBASE_USER_EMAIL=user@example.com
FIREBASE_USER_PASSWORD=your-password-here
# Optional
PORT=3003Note: You can find these values in your Firebase Console under Project Settings.
Choose your preferred transport protocol:
# stdio (default - for MCP clients like Claude Desktop)
bun run stdio
# HTTP Streamable
bun run http
# Server-Sent Events (SSE)
bun run sseFireMCP provides five Firestore operations as MCP tools:
Retrieve a document from Firestore by its path.
Input:
{
"path": "users/userId"
}Output:
{
"exists": true,
"data": { "name": "John Doe", "email": "john@example.com" }
}Create or overwrite a document in Firestore.
Input:
{
"path": "users/userId",
"data": { "name": "Jane Doe", "email": "jane@example.com" },
"merge": false
}Output:
{
"success": true
}Add a new document to a collection with an auto-generated ID.
Input:
{
"collection": "users",
"data": { "name": "Bob Smith", "email": "bob@example.com" }
}Output:
{
"id": "auto-generated-id"
}Delete a document from Firestore.
Input:
{
"path": "users/userId"
}Output:
{
"success": true
}Query a Firestore collection with filters and limits.
Input:
{
"collection": "users",
"filters": [
{
"field": "age",
"operator": ">",
"value": 18
},
{
"field": "active",
"operator": "==",
"value": true
}
],
"limit": 10
}Output:
{
"documents": [
{ "id": "doc1", "name": "Alice", "age": 25, "active": true },
{ "id": "doc2", "name": "Bob", "age": 30, "active": true }
]
}Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):
{
"mcpServers": {
"firestore": {
"command": "bun",
"args": ["run", "/path/to/firemcp/index.ts"],
"env": {
"FIREBASE_API_KEY": "your-api-key",
"FIREBASE_PROJECT_ID": "your-project-id",
"FIREBASE_USER_EMAIL": "user@example.com",
"FIREBASE_USER_PASSWORD": "your-password"
}
}
}
}For debugging and testing:
bun run inspectbun run start- Start with default transport (stdio)bun run dev- Same as startbun run http- Start with HTTP Streamable transportbun run sse- Start with SSE transportbun run stdio- Start with stdio transportbun run inspect- Launch MCP Inspector for debuggingbun run kill- Kill any process running on port 3003
- Create a new file in
src/tools/ - Define input/output schemas using Zod
- Implement the tool function
- Register the tool in
src/server.ts
Example:
import { z } from "zod";
import { getFirestoreInstance } from "../firestore.js";
const inputSchemaObject = z.object({
// Define your input schema
});
const outputSchemaObject = z.object({
// Define your output schema
});
export const inputSchema = inputSchemaObject.shape;
export const outputSchema = outputSchemaObject.shape;
export const myTool = async (input: z.infer<typeof inputSchemaObject>) => {
const db = await getFirestoreInstance();
// Implement your tool logic
return {
content: [{ type: 'text' as const, text: JSON.stringify(result) }],
structuredContent: result
};
};Since FireMCP uses the Client SDK, you must configure Firestore Security Rules. Example rules:
rules_version = '2';
service cloud.firestore {
match /databases/{database}/documents {
// Allow authenticated users to read/write their own data
match /users/{userId} {
allow read, write: if request.auth != null && request.auth.uid == userId;
}
// Allow authenticated users to read public collections
match /public/{document=**} {
allow read: if request.auth != null;
}
}
}This project is licensed under the MIT License - see the LICENSE file for details.
- Built with the Model Context Protocol SDK
- Powered by Firebase
- Runtime by Bun
If you have any questions or run into issues, please open an issue on GitHub.