Advanced Features

Since v2.0.0

Firestore MCP server

Fuego's MCP server gives your AI tool 50 tools for Cloud Firestore, the Realtime Database, Firebase Authentication and Cloud Storage, using the projects you already added to Fuego. Let the AI tool start it with fuego mcp serve, or connect to the server built into the app, and decide which projects it may reach and whether it may write.

What it is

The Model Context Protocol (MCP) is the standard way for AI tools such as Claude Code, Claude Desktop, Cursor and VS Code to use outside tools. Fuego includes an MCP server, so the AI tool you already work with can read and change your Firestore data through Fuego: list collections, infer a schema, run a query, look up an Authentication user, download a Storage file, export a collection.

It works with the projects you already added to Fuego. There are no credentials to copy into the AI tool, and the rules you set in Fuego decide which projects it may reach, whether it may write and whether personal data is masked before the AI tool sees it.

There are two ways to connect:

Your AI tool starts FuegoThe server built into the app
HowThe AI tool runs fuego mcp serve when it needs itTurn it on in Settings ▸ AI ▸ MCP Server and give the AI tool its URL
Fuego app openNot neededNeeded
Access tokenNot neededOptional
Best forClaude Code, Claude Desktop, Cursor, VS Code and most other toolsTools that connect by URL, other devices on your network

Both offer the same tools and follow the same rules from Settings ▸ AI ▸ Data access.

What you need

  • Fuego, on any plan: the MCP server is included in the free plan too, for the one project the free plan allows (emulators need a paid plan). Anonymize tool results needs a paid plan (Individual or Team, including the 14-day free trial, or the Lifetime Deal). See pricing.
  • The fuego command, for the first way. Install it from Help ▸ Install Command Line Tool… in Fuego (with the custom title bar, the menus are in the menu button of the toolbar); fuego mcp serve needs Fuego 2.2 or later. See Command line.
  • At least one project added to Fuego. Your AI tool works on the projects you added to Fuego.

An unrelated open-source Firestore tool also installs a command called fuego: the command your AI tool runs must be the one installed by Fuego from Help ▸ Install Command Line Tool….

TIP

Run fuego mcp info in a terminal. It prints, ready to paste, the Claude Code command and the mcpServers JSON for both ways of connecting, and, when no access token is required, the Codex command for the server built into the app. The VS Code entries, the --scope user variant and the install links are on this page only.

Let your AI tool start Fuego

In this setup the AI tool runs fuego mcp serve on demand and talks to it directly (what MCP directories call a stdio server). Nothing needs to be open or switched on in Fuego, and no token is involved: only the AI tool that started the command can talk to it. The Enable MCP server switch in Settings concerns the server built into the app, not this one.

Claude Code

claude mcp add fuego -- fuego mcp serve

Claude Code registers the server for the current folder. To make it available in every folder:

claude mcp add --scope user fuego -- fuego mcp serve

Claude Desktop

In Claude Desktop, open Settings ▸ Developer ▸ Edit Config and add Fuego to the configuration file, then quit and reopen Claude Desktop:

{
  "mcpServers": {
    "fuego": {
      "command": "fuego",
      "args": ["mcp", "serve"]
    }
  }
}

If the file already has an mcpServers section, add the "fuego" entry inside it. Claude Desktop does not read the PATH set in your terminal, so if it reports that it cannot start the server, replace "fuego" with the full path of the app. On macOS that is:

/Applications/Fuego.app/Contents/MacOS/Fuego

On Windows and Linux, fuego cli status shows it on the Executable line, or look in the folder where you installed Fuego. On Windows, write every backslash of the path twice inside the JSON ("C:\\...\\Fuego.exe") or use forward slashes ("C:/.../Fuego.exe"): a single backslash makes the file invalid JSON, and Claude Desktop cannot read it.

Cursor

Add Fuego to Cursor opens Cursor with the server ready to install. To do it by hand, add the same entry to ~/.cursor/mcp.json (every project) or .cursor/mcp.json (one project):

{
  "mcpServers": {
    "fuego": {
      "command": "fuego",
      "args": ["mcp", "serve"]
    }
  }
}

VS Code

Add Fuego to VS Code opens VS Code with the server ready to install. To share it with a project, create .vscode/mcp.json:

{
  "servers": {
    "fuego": {
      "type": "stdio",
      "command": "fuego",
      "args": ["mcp", "serve"]
    }
  }
}

VS Code uses the key servers, not mcpServers.

Other MCP clients

Any tool that can start a local MCP server works the same way: the command is fuego and the arguments are mcp and serve.

For a tool that only connects by URL, fuego mcp serve --http serves the same tools on this computer only, with an access token created for that run and printed when it starts. Add --port 43111 to choose the port; otherwise a free one is picked. Fuego’s own MCP settings are not changed.

Connect to the server built into the app

The app can also run the server itself, at a URL on your computer. It runs only while Fuego is open.

  1. Open Settings ▸ AI ▸ MCP Server and turn on Enable MCP server. Start automatically when Fuego launches brings it back every time you open Fuego.
  2. Click Show connection info and copy the snippet for your AI tool: Claude Code CLI, JSON configuration (mcpServers) or Codex CLI.

MCP Server settings in Fuego

The default address is http://127.0.0.1:43110/mcp, which only programs on this computer can reach. For Claude Code:

claude mcp add --transport http fuego http://127.0.0.1:43110/mcp

For tools configured with a JSON file:

{
  "mcpServers": {
    "fuego": {
      "type": "http",
      "url": "http://127.0.0.1:43110/mcp"
    }
  }
}

You can also start and stop the server from the AI menu, the command palette, the MCP server card on the Home page, or a terminal with fuego mcp start and fuego mcp stop. fuego mcp info prints the URL, the token and the snippets even while Fuego is closed. If a token is required but has not been created yet, it says so instead: Fuego creates the token when it starts the server.

The same panel has the connection options:

  • Port: 43110 by default.
  • Require bearer token: every request must carry the token. The snippets shown by Fuego and by fuego mcp info include it. Regenerate replaces it; the new token is shown once.
  • Serve over HTTPS: with a certificate Fuego creates and renews, or with your own. Some AI tools only accept https:// addresses; a certificate made by Fuego may need to be trusted in the AI tool first.
  • Advertise on the local network (mDNS): lets other devices on your network find and reach the server. The token is then always required, and macOS asks for permission to use the local network.
  • Compatibility mode for older clients: see Troubleshooting.

Choose what the AI can reach

The rules live in Settings ▸ AI ▸ Data access and apply to both connection types and to Fuego’s own AI assistant:

  • Which projects and databases: All projects, or a list you pick. With nothing picked, every request is refused.

  • Which Realtime Databases: with All projects, every Realtime Database of your projects. With a list, only the databases you pick under Realtime Databases: all the Realtime Databases of a project, or some of them. A list saved before Fuego could reach Realtime Databases reaches none of them until you pick some, and while none is picked the AI tool is not offered the Realtime Database tools at all. Picking a Realtime Database opens that database only: the project’s Firestore databases, Authentication users and Storage files stay closed unless you pick a Firestore database of the project too.

  • Write access:

    • Do not allow writes: only the read tools are offered.
    • Allow writes only on non-production projects (the default): every tool is offered, but projects tagged as production stay read-only.
    • Allow write operations: writes on every project in scope.

    A project or database set to read-only in Fuego can never be written, whatever you choose here. Exports count as writes: they create files on your computer and run as background jobs. An AI tool connected over MCP cannot choose where they go: every export is written to the Export folder set in Settings ▸ AI, under a file name the tool may pick, and never replaces a file already there.

    An AI tool never replaces or deletes a whole Realtime Database, never runs a Realtime Database job that would, and never makes a Realtime Database change that Fuego would ask you to confirm by typing its key: removing 1,000 children or more, 10 nodes or more, or more than Fuego can count first. Such a change is refused, and you make it in Fuego. Fuego’s own AI assistant can make these large changes: its card asks you for the key, its Approve all gives that key for you once you type the code shown next to it, and fuego ai --approve-tools gives it too.

    An AI tool connected over MCP never names a file of your computer to a Realtime Database job either: it cannot import a file, and its Realtime Database exports go into the Export folder like the others. Fuego’s own AI assistant can export to a file you name and import a JSON or JSON Lines file, after you confirm: its card names the file.

  • Anonymize tool results (in the MCP Server panel, paid plans): documents, Authentication users and Realtime Database values are masked according to your anonymization settings before the AI tool receives them; the keys of a Realtime Database stay as they are. For an aggregation, Fuego decides what to mask in each answer: its groups, the total of a grouped sum or average, and the results it takes for the value of a single child are masked (see rtdb_aggregate below). An AI tool can ask for masking on a single request, but can never turn it off.

The server built into the app applies a change right away. When your AI tool starts Fuego, the settings are read at that moment: restart or reconnect the server in your AI tool after a change.

WARNING

Whatever a tool returns, documents included, goes to your AI tool and from there to the AI model it uses. Limit the projects in Data access, or turn on Anonymize tool results, for data that must not leave your computer.

The 50 tools

AreaRead (27)Write (23)
Firestore: collections, schemas, queries, counts, sums and averagesfirestore_list_collections, firestore_get_schema, firestore_query, firestore_get_document, firestore_count, firestore_sum, firestore_avgfirestore_create_document, firestore_update_document, firestore_delete_document
Realtime Database: databases, nodes, queries, aggregationsrtdb_list_instances, rtdb_get, rtdb_query, rtdb_aggregatertdb_set, rtdb_update, rtdb_push, rtdb_delete
Indexes: list them, analyze a query or find duplicatesindexes_list, indexes_suggestindexes_create, indexes_update, indexes_delete
Authentication: users, tenants and custom claimsauth_list_users, auth_get_user, auth_count_users, auth_get_tenantsauth_update_user, auth_delete_user
Storage: buckets, files and their metadatastorage_list_buckets, storage_list_files, storage_get_file, storage_download_filestorage_update_file, storage_delete_files
Background jobs: copy, move, rename, bulk update and delete, point-in-time restorejobs_list, jobs_getjobs_submit, jobs_stop, rtdb_submit_job
Exports: collections, users, Storage files and Realtime Database nodes to a file on your computerexport_collection, export_users, export_storage, export_rtdb
Projects: the projects and databases added to Fuegoprojects_list, projects_get
Bookmarks and saved queriesbookmarks_list, queries_listbookmarks_save, bookmarks_delete

Fuego’s query builder features are available to the AI tool too: firestore_query accepts filters, sorting and cursors, and pipeline stages on Enterprise edition databases. User exports never include password hashes. jobs_submit never reads or writes a file on your computer: exports go through the export tools, and imports and uploads are done in Fuego itself.

Since Fuego 2.3.0, the server has Realtime Database tools, offered only when Data access reaches a Realtime Database. They name a database by its name, as rtdb_list_instances lists it, and exchange values as JSON text, so large numbers keep every digit. rtdb_query takes one ordering with its bounds and limit, as a query on the Realtime Database page does; ranges and limitToLast need a paid plan, as in the app. The AI tool can also run as an unauthenticated user or as a user of your app, to see what your security rules let them read or write.

rtdb_aggregate counts the children of a node, or adds up, averages or finds the minimum or maximum of a child, optionally grouped by a child or compared with the previous period: the aggregations of the Realtime Database widgets of a dashboard, plus the minimum and the maximum. It makes one read of at most 9,999 children, gives exact decimal results and at most 100 groups. It always runs as admin, so your security rules do not apply to it.

With Anonymize tool results on, Fuego decides what to mask in each answer. The groups are masked, as they are values of your data; a group that holds an object or a list may read only ****. A minimum or a maximum is the value of one child, and is masked as that child’s value would be, for the whole result and for each group alike. A sum or an average is masked when it could be the value of one child too: when it covers a single child (masked as that child’s value would be), when the average of its numbers is a whole number that looks like a card number (as it is when every child holds the same card number), or, whatever its numbers, when the child is a personal field such as phone. The previous period of a sum or an average is checked the same way, and so is each group of a grouped one. The total of a grouped sum or average, and its previous period, are always masked: next to the groups, they could reveal the value of a single child. Counts are never masked, and most sums and averages of amounts, such as order totals, are not either. A sum, average, minimum or maximum of children that are plain numbers with no field name (such as a list of scores by user), and a grouping by such values, are refused: those numbers cannot be told apart from card or phone numbers. The AI tool reads them with rtdb_get or rtdb_query instead, which mask each value.

export_rtdb exports a node, some of its children or the children a query matches, as JSON, JSON Lines, CSV or YAML, into the Export folder. rtdb_submit_job runs the Realtime Database’s background jobs: delete, copy, move or rename a node (to another path, database or project) and update every matching child, as the activities of the Realtime Database page do, with the same plan limits. Fuego first works out what a job would remove and applies the rule above: a job that would need its key typed is refused. Both run in the background; jobs_list and jobs_get follow them, and show a Realtime Database job only when Data access reaches its databases, the destination of a copy or a move included.

When your AI tool starts Fuego with fuego mcp serve, its jobs run inside that command, which waits for the jobs still running when the AI tool ends the session. The MCP specification has the AI tool stop the command a few seconds after it ends the session, though, and a job still running then stops there, part done: let the AI tool follow it with jobs_get until it has ended. Jobs started through the server built into the app run in the app.

The server also offers two ready-made prompts, Analyze a Firestore collection and Write a Fuego query, and lets the AI tool read the list of projects, the collections of a database, the schema of a collection, the keys at the top of a Realtime Database and the shape of a Realtime Database query.

Example prompts

Once connected, ask in your own words; the AI tool picks the tools:

  • “List my Firebase projects in Fuego and their databases.”
  • “What does a document in orders look like in shop-staging? Which fields are optional?”
  • “Find the orders with status failed from the last 7 days and add up their total.”
  • “How many Authentication users does shop-prod have? Show me the custom claims of ada@example.com.”
  • “Do I have duplicate composite indexes in the default database?”
  • “Export customers from shop-staging to CSV with only email and country.”
  • “List the files under invoices/2026/ in the default bucket and download the newest one.”
  • “In the Realtime Database of shop-staging, show me what is stored under /carts/ada.”
  • “Export /orders from the Realtime Database of shop-staging to JSON Lines.”

Troubleshooting

The server stops as soon as the AI tool starts it. Run the command in a terminal to see why:

fuego mcp serve

Exit code 5 means the license or plan does not allow it: an expired license, or the limit of computers for your license. On the free plan, a tool that names a project beyond the plan’s one project, or an emulator, answers that the plan does not allow it. A server that starts correctly waits for an AI tool to talk to it; press Ctrl+C to stop it.

command not found, or the AI tool cannot start fuego. Install the command from Help ▸ Install Command Line Tool…. AI tools opened from the Dock or the Start menu do not read the PATH set in your terminal, so if yours still cannot start fuego, use the full path of the app as the command instead. On macOS:

/Applications/Fuego.app/Contents/MacOS/Fuego

On Windows and Linux, fuego cli status shows it on the Executable line, or look in the folder where you installed Fuego. In a JSON config file on Windows, write every backslash of the path twice or use forward slashes. If your terminal does not recognise fuego cli status, the fuego it finds is a different program: see What you need.

An older AI tool refuses to connect to the URL. Each AI tool gets the newest protocol version it supports, and older ones keep working. If one still refuses, turn on Compatibility mode for older clients in Settings ▸ AI ▸ MCP Server and reconnect. It also applies to fuego mcp serve --http.

Write tools are missing, or a write is refused. Data access is set to Do not allow writes, the project is tagged as production, or the project or database is read-only in Fuego. A project left out of Data access is refused for reads too.

The Realtime Database tools are missing, or a Realtime Database is “not in the allowed databases” (mcp_rtdb_instance_not_allowed). Data access is set to a list that picks no Realtime Database, or not this one: pick it under Realtime Databases. A list saved before Fuego could reach Realtime Databases picks none.

A Realtime Database change “deletes too much to be made from an MCP client” (mcp_rtdb_destructive_refused). The change would remove 1,000 children or more, 10 nodes or more, or more than Fuego could count first: make it in Fuego, which asks you to type its key. The same holds for a Realtime Database job, and for any job that would delete or replace a whole database.

A Realtime Database import, or an export to a chosen file, is “not available to MCP clients” (mcp_rtdb_local_file_refused). An AI tool connected over MCP never names a file of your computer: import the file in Fuego, and let the export go into the Export folder (it may name the file there).

The port is in use. With Fuego closed, fuego mcp status tells you whether another program already listens on the configured port. Pick another Port in the panel.