Ask the AI assistant
- Open a collection, or any tab of the project you want to ask about.
- Choose Tools, AI assistant, press Ctrl+I (⌘I on a Mac), or type "assistant" in the command palette.
- The first time, Firetool explains what the assistant sends and to whom. Choose I agree to go on, or Not now.
- Paste your Anthropic API key. Create one in the Anthropic Console, under API keys. Firetool keeps it encrypted on your computer and sends it only to Anthropic.
- Ask in your own words, for example "How many orders are still open?" or "Which fields do users have, and how often?". Pick one of the examples to start.
While it works, the panel lists each step: what it read (a count, a collection's fields, a query) and how many documents. The answer follows. Ask a follow-up in the same conversation, or choose New to start again. The assistant works on one project at a time: open it from another project's tab and it starts a new conversation for that project. Stop ends an answer after the current step.
In Settings, General, AI assistant you can change the model (a larger one answers harder questions, a smaller one is faster and costs less on your Anthropic bill), turn on Hide personal data, remove the key, or withdraw your agreement.
Each document the assistant reads is a Firestore read, billed like any other. It keeps reads small: counts for how many, and a limit on queries. The assistant itself is billed by Anthropic to your account.
Changes it suggests
The assistant can't change anything. When you ask for a change, it proposes one: what to create, replace, update or delete, and why.
- Choose Review the proposed change under its answer, or the AI: 1 to review button beside the edition badge.
- AI assistant changes shows the reason and each document as it is now beside the change.
- Approve and write writes all the changes together, or none. Reject writes nothing.
Approving goes through the same checks as any change in Firetool: your role, read-only and production projects (a production project asks for its ID), and the audit log. If a document changed after the assistant proposed the change, nothing is written, and you can ask it to propose again. A proposal you don't decide on within a day expires.
Writing an approved change needs Pro. Asking questions doesn't.
What is sent
To answer, Firetool sends to Anthropic, from your computer: your question, the earlier questions and answers of the conversation, and what the assistant reads to answer it, for the open tab's project only. With Hide personal data on, email addresses and phone numbers in what it reads are masked first. Nothing goes to us. What Anthropic does with it is covered by your agreement with Anthropic; see the privacy policy.
Connect your own assistant (MCP)
Assistants and editors that support the Model Context Protocol (MCP) can use Firetool as a server: the firetool mcp command, which comes with the app.
- In Firetool, open Settings, General, AI assistants (MCP) and turn on Let AI assistants read through Firetool.
- Under Projects it may read, tick the projects the assistant may use. Nothing is allowed until you tick it.
- Choose Copy the entry and add it to your assistant's MCP settings. It looks like this, with the path of the firetool command on your computer, which Settings shows:
{
"mcpServers": {
"firetool": {
"command": "<the firetool command Settings shows>",
"args": ["mcp"]
}
}
}
To let the assistant suggest changes too, turn on Let assistants propose changes. Its proposals wait in AI assistant changes exactly like the built-in assistant's, and only someone in Firetool's window can approve them.
The assistant uses the accounts already in Firetool, so it needs no keys of its own. Turning the switch off stops it at its next call.
What an assistant can do
- List the projects you allowed, the collections of a database, and the subcollections of a document.
- Get a document; query a collection with filters, sorting and a limit; count documents.
- Describe a collection's fields from a sample: their types and how often each appears.
- Read the composite indexes, the Firestore or Storage security rules, an Authentication user by email, phone or UID, and the files in a Storage folder.
- Propose changes to documents, and check what became of a proposal.
The audit log
Every call is in the audit log on your computer: assistant.read for the built-in assistant, mcp.read for one you connect, and .denied for refused calls, with the project and what was read. Proposals are recorded as assistant.propose or mcp.propose, rejections as mcp.reject, and an approved change is marked as the assistant's. View its calls in Settings opens the audit log filtered on them.
Questions
Which AI provider does the built-in assistant use?
Anthropic, with your own API key. Choose the model in Settings, General, AI assistant. An assistant you connect through MCP uses whatever provider it uses.
Can the assistant delete or change my data?
No. It can only propose changes. Nothing is written until someone approves the proposal in Firetool's window, and approving follows your roles, read-only and production rules.
Does my data go to Firetool's servers?
No. The built-in assistant sends from your computer straight to Anthropic, and an assistant you connect gets the data on your computer. Nothing goes to us.
Can I keep personal data out of it?
Turn on Hide personal data: email addresses and phone numbers are masked before they leave Firetool. You can also allow only some projects for MCP, and the built-in assistant only ever works on the project you opened it from.
Related
- AI assistant and MCP: what it is and how it keeps your data safe.
- Roles and production safety: the checks every approved change goes through.
- The firetool command: the same Firetool from a terminal.