Codex CLI for Firebase Development: MCP Server, Agent Skills, and Full-Stack Workflows

Updated

codex-clifirebasemcpagent-skillsfirestoreauthenticationcrashlyticscloud-functionsfull-stack

Codex CLI for Firebase Development: MCP Server, Agent Skills, and Full-Stack Workflows

Firebase powers over 3.5 million active applications1, yet until recently, developing against its services from an AI coding agent meant blind prompting — the agent hallucinating collection paths, guessing security rule syntax, and fabricating Crashlytics issue IDs. That changed when the Firebase MCP server graduated to general availability in October 20252 and the Firebase agent skills catalogue landed in early 20263. Together, they give Codex CLI verified, tool-backed access to the entire Firebase platform — from Firestore queries to push notification delivery — without leaving the terminal.

This article covers the official Firebase MCP server (built into Firebase CLI v15.18.04), the Firebase agent skills system, and the community gannonh/firebase-mcp server5, then walks through four practical workflow patterns for full-stack Firebase development with Codex CLI.

The Firebase MCP Server

The Firebase MCP server ships inside firebase-tools itself — no separate package to install2. It exposes 50+ tools across twelve Firebase service categories via stdio transport6:

Category Key Tools Purpose
Core firebase_get_project, firebase_list_apps, firebase_get_sdk_config Project discovery and configuration
Firestore firestore_get_documents, firestore_query_collection, firestore_list_collections, firestore_delete_document Document CRUD and collection enumeration
Authentication auth_get_users, auth_update_user, auth_set_sms_region_policy User management and custom claims
Crashlytics crashlytics_get_issue, crashlytics_list_events, crashlytics_get_report, crashlytics_create_note Crash investigation and annotation
Cloud Functions functions_list_functions, functions_get_logs Function listing and log retrieval
Cloud Messaging messaging_send_message Push notification delivery
App Hosting apphosting_list_backends, apphosting_fetch_logs Backend management and log access
SQL Connect dataconnect_build, dataconnect_execute GraphQL operations against Cloud SQL
Storage storage_get_object_download_url Download URL generation

Additionally, the server provides security rule tools — firebase_validate_security_rules and firebase_get_security_rules — that let the agent read, generate, and validate rules for Firestore, Cloud Storage, and Realtime Database without guessing syntax6.

Configuring the Firebase MCP Server for Codex CLI

Add the server to ~/.codex/config.toml:

[mcp_servers.firebase]
command = "npx"
args = ["-y", "firebase-tools@latest", "mcp"]

To scope the server to a specific project directory and limit tool categories:

[mcp_servers.firebase]
command = "npx"
args = ["-y", "firebase-tools@latest", "mcp", "--dir", "/home/dev/my-app", "--only", "firestore,auth,crashlytics"]

The --only flag is worth using. Exposing all 50+ tools consumes context tokens; filtering to the services you actually use keeps the tool manifest lean6.

Since v15.16.0, the Firebase MCP server also supports SSE mode with a customisable port4, useful when sharing a single server instance across multiple agent sessions:

npx firebase-tools@latest mcp --transport sse --port 9199

For Codex CLI’s streamable HTTP transport:

[mcp_servers.firebase-sse]
url = "http://localhost:9199/sse"

Authentication

The server uses your Firebase CLI credentials. Run firebase login once, and the MCP server inherits the session — no separate API key management required6. For CI environments, set GOOGLE_APPLICATION_CREDENTIALS to a service account key path and the server picks up Application Default Credentials automatically.

Firebase Agent Skills

Firebase agent skills are portable, self-contained knowledge modules that provide Firebase-specific instructions, best practices, and automation scripts3. Where the MCP server provides tool access (read this document, query that collection), skills provide domain knowledge (how to structure Firestore for offline sync, when to use custom claims versus security rules).

Install skills for Codex CLI:

npx skills add firebase/agent-skills --agent=codex

Key skills available:

Skill Focus
firebase-auth-basics Sign-in flows, custom claims, multi-factor authentication
firebase-firestore-standard Data modelling, indexing, offline persistence, security rules
firebase-app-hosting-basics Framework deployment (Next.js 16, Angular, SvelteKit)
firebase-ai-logic-basics Gemini API integration via Firebase AI Logic
developing-genkit-js Building AI-powered applications with Genkit in Node.js

Skills use progressive disclosure — the agent initially scans brief metadata and loads detailed instructions only when the task matches3. This keeps token costs low compared to dumping entire documentation sets into the system prompt.

The Community Firebase MCP Server

For teams needing capabilities beyond the official server, gannonh/firebase-mcp5 provides an alternative with HTTP transport support and file upload tools:

  • storage_upload — upload files from text, base64 content, or local file paths
  • storage_upload_from_url — import files directly from external URLs with automatic content type detection

Configure it alongside the official server:

[mcp_servers.firebase-community]
command = "npx"
args = ["-y", "@gannonh/firebase-mcp"]

[mcp_servers.firebase-community.env]
FIREBASE_PROJECT_ID = "my-project-id"
SERVICE_ACCOUNT_KEY_PATH = "/path/to/service-account.json"

⚠️ The community server requires explicit service account configuration — it does not inherit Firebase CLI credentials.

AGENTS.md for Firebase Projects

Create AGENTS.md at the project root to encode Firebase-specific conventions:

# AGENTS.md — Firebase Project Conventions

## Stack
- Firebase CLI v15.18.0, Node.js 22 LTS
- Firestore in Native mode (NOT Datastore mode)
- Cloud Functions 2nd gen (Node.js 22 runtime)
- Firebase Authentication with Identity Platform

## Rules
- All Firestore security rules MUST use `rules_version = '2'`
- Never use `allow read, write: if true` in any environment
- Custom claims are for role-based access; do not store user data in claims
- Cloud Functions must use `onRequest` (2nd gen) not `functions.https.onRequest` (1st gen)
- Firestore composite indexes must be defined in `firestore.indexes.json`, not created manually

## Anti-hallucination
- Collection paths: verify with `firestore_list_collections` before assuming a path exists
- Function names: verify with `functions_list_functions` before referencing
- Security rules: always validate with `firebase_validate_security_rules` after generation
- SDK config: retrieve with `firebase_get_sdk_config`, never hardcode project IDs

## Testing
- Use Firebase Emulator Suite for all local development
- Emulator ports: Auth 9099, Firestore 8080, Functions 5001, Storage 9199

Workflow Patterns

Pattern 1: Crash Investigation with Crashlytics MCP

The most immediate value of the Firebase MCP server is crash triage. Instead of switching to the Firebase console, the agent queries Crashlytics directly:

flowchart LR
    A[Identify crash spike] --> B[crashlytics_get_report]
    B --> C[crashlytics_get_issue top_issue_id]
    C --> D[crashlytics_list_events issue_id]
    D --> E[Agent analyses stack traces]
    E --> F[Generate fix + tests]
    F --> G[crashlytics_create_note documenting fix]

Prompt:

The crash rate spiked overnight. Use the Crashlytics MCP tools to pull the
top issues, examine the stack traces, identify the root cause, write a fix,
and annotate the Crashlytics issue with what you found.

The agent calls crashlytics_get_report for numerical overview, drills into the top issue with crashlytics_get_issue, examines individual events via crashlytics_list_events, proposes a fix, and documents its findings with crashlytics_create_note — all without leaving the terminal6.

Pattern 2: Security Rules Generation and Validation Loop

Firestore security rules are notoriously error-prone. The MCP server enables a generate-validate-fix loop:

flowchart TD
    A[Read existing rules] --> B[firebase_get_security_rules]
    B --> C[Agent generates new rules]
    C --> D[firebase_validate_security_rules]
    D -->|errors| C
    D -->|valid| E[Write rules to firestore.rules]
    E --> F[firebase deploy --only firestore:rules]

The agent retrieves the current rules, understands the data model via firestore_list_collections and firestore_get_documents, generates updated rules, and validates them before writing to disk. The validation loop catches syntax errors and type mismatches that would otherwise only surface at deploy time6.

Pattern 3: Full-Stack Feature Development

Combining MCP tools with agent skills enables end-to-end feature development:

Add a user profile page. Use Firestore for storage, Firebase Auth for the
session, and Cloud Storage for the avatar. Generate security rules that
allow users to read/write only their own profile and upload images under 5MB.
Validate everything before committing.

The agent:

  1. Queries firestore_list_collections to understand the existing schema
  2. Retrieves SDK config via firebase_get_sdk_config
  3. Scaffolds the feature using knowledge from firebase-firestore-standard and firebase-auth-basics skills
  4. Generates Firestore and Storage security rules
  5. Validates rules with firebase_validate_security_rules
  6. Writes the implementation and associated tests

Pattern 4: Batch Audit with codex exec

For multi-project Firebase estates, codex exec enables batch operations:

for project_dir in projects/*/; do
  codex exec \
    --model gpt-5.4-mini \
    --approval-mode full-auto \
    "Audit the Firestore security rules in this project. Flag any rule that \
     grants public write access or allows reads without authentication. \
     Output findings as JSON." \
    < "$project_dir/firestore.rules"
done

This pattern works because the Firebase MCP server’s --dir flag scopes tool access to each project directory6.

Model Selection

Task Recommended Model Rationale
Security rules generation/audit o3 Rules require precise logic; reasoning tokens catch edge cases
Crash investigation and triage gpt-5.5 Complex stack trace analysis benefits from frontier reasoning7
CRUD scaffolding and boilerplate gpt-5.4-mini Routine generation; cost-effective at scale
Batch rule audits via codex exec gpt-5.4-mini High volume, structured output

Composing MCP Servers

The Firebase MCP server composes well with other servers for full-stack workflows:

[mcp_servers.firebase]
command = "npx"
args = ["-y", "firebase-tools@latest", "mcp", "--only", "firestore,auth,crashlytics"]

[mcp_servers.github]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]

[mcp_servers.filesystem]
command = "npx"
args = ["-y", "@anthropic-ai/mcp-server-filesystem", "/home/dev/my-app"]

This gives the agent Firebase project access, GitHub issue/PR context, and local filesystem operations — the three layers needed for most full-stack development workflows.

Sandbox Considerations

  • Network access required: The Firebase MCP server calls Firebase APIs over HTTPS. Codex CLI’s sandbox must allow outbound connections to *.googleapis.com and *.firebaseio.com8
  • Emulator alternative: For offline development, start the Firebase Emulator Suite and point the MCP server at local emulators — no network egress needed
  • Credential exposure: The MCP server inherits your Firebase CLI session. In full-auto mode, the agent can modify production data. Use suggest or auto-edit approval modes for production projects8
  • npx cold starts: First invocation downloads firebase-tools. Pin a version (firebase-tools@15.18.0) to avoid unexpected updates mid-session

Limitations

  • Training data lag: Codex models may not know about Firebase CLI v15.18.0 features (SQL Connect search indexes, Crashlytics report tools) — the MCP server bridges this gap by providing live tool access7
  • No emulator control: The MCP server cannot start or stop the Firebase Emulator Suite; you must manage emulators separately
  • Storage tool gaps: The official server only provides storage_get_object_download_url — no upload, delete, or list operations. Use gannonh/firebase-mcp for upload workflows5
  • SQL Connect maturity: The dataconnect_* tools reflect Firebase’s renamed Data Connect → SQL Connect service; the API surface is still evolving4
  • Token budget: With 50+ tools registered, the Firebase MCP server consumes significant context. Use --only to filter to relevant service categories

Citations

  1. Firebase Official Site — “Used by millions of businesses” ↩

  2. Firebase Blog — “It’s graduation day for the Firebase MCP Server” ↩ ↩2

  3. Firebase Docs — Agent Skills ↩ ↩2 ↩3

  4. Firebase CLI Release Notes — v15.16.0–v15.18.0 ↩ ↩2 ↩3

  5. gannonh/firebase-mcp — GitHub ↩ ↩2 ↩3

  6. Firebase Docs — MCP Server ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  7. OpenAI — Codex CLI Models ↩ ↩2

  8. OpenAI — Codex Permissions and Sandbox ↩ ↩2