Connect Claude and ChatGPT to Filetable (MCP)

Why connect an assistant to Filetable

An AI assistant is good at understanding a question and writing an answer, but it has never seen your invoices, receipts, contracts or bank statements. Filetable has: it is the finance and admin specialist that reads every document you send it, scans and photos included, extracts the figures into linked tables, files it in the right folder and matches it to the others: an invoice to its supplier and its lines, a bank line to the invoice it pays. The connector lets your assistant work on all of that. You ask where you already work, and the answer comes from your own figures instead of a guess.

• Answers from your figures. A total is computed by Filetable over every matching row, not added up in the model's head.
• The document behind every value. The assistant finds a contract by its content, reads what Filetable extracted from it, and can read the document's own text to check.
• The admin done, not just explained. Record a payment, link a bank line to its invoice, correct a value, file a document: the assistant does it with your playbooks and under the same rules as you in the app.
• Your accountant stays your accountant. Filetable prepares and checks the data; bookkeeping, tax returns and year-end closing stay with your accounting software and your accountant, who can work from that data.
• One conversation for several tools. The assistant can combine Filetable with your other connectors (your mail, your calendar) in the same request.

Before you start

• A Filetable account, member of at least one workspace. The connector is open in every workspace; the MCP tab of Settings lists your connections.
• An app that accepts a remote MCP server: Claude (web, desktop, mobile), Claude Code, or ChatGPT depending on your plan (see below).
• Two minutes. There is nothing to install and no token to copy for Claude and ChatGPT.

In Claude (web, desktop and mobile)

On the Free, Pro and Max plans you add the connector yourself. The Free plan takes one custom connector. Add it from the web or the desktop app: it is then available on mobile the next time you sign in.

Claude Team and Enterprise

In an organization, an owner first adds the connector for everyone, in Organization settings → Connectors, with the same URL. Each member then finds it in Customize → Connectors, labelled Custom, and clicks Connect: the approval in Filetable stays personal, so each person lends only their own rights.

Claude Code

In a terminal, one command adds Filetable; the first use of /mcp opens the same approval page in your browser.

claude mcp add --transport http filetable https://filetable.com/api/mcp# then, inside Claude Code:/mcp

For a script, a CI job or a client without a browser sign-in, create a personal token instead in Settings → MCP → New connection: you choose its name, expiry and workspaces, and Filetable shows the token once with the command ready to copy. Only its hash is kept; a lost token is revoked and replaced.

claude mcp add --transport http filetable https://filetable.com/api/mcp \  --header "Authorization: Bearer ftmcp_…"

In ChatGPT

ChatGPT takes a custom MCP server from its settings. OpenAI renamed that part several times in 2026 (connectors, apps, developer mode, plugins), but the flow stays the same: declare a server, then approve it in Filetable.

What you grant when you approve

The only moment any access is given is Filetable's approval screen. The assistant sees none of it: it receives a one-hour key, renewed while the connection lives, worth exactly what you ticked. Read only, the write tools are not even offered to it.

You also choose the workspaces: all of them, including those you join later, or only the ones you tick. Whatever you tick, the connection never goes past your role: a read-only member of a workspace does not write there, and a table or folder closed to you is closed to the assistant too.

The assistant's 40 tools

A connected assistant sees a list of tools, each with a description telling it when to call it. You do not have to name them: ask in plain words and it picks. Here they are, grouped by what they are for.

A few limits, on purpose: query_table returns at most 100 rows a call (so a total goes through aggregate_table, which Filetable computes over every row); get_row shows the first 3 linked rows of each relation and counts the rest, which get_row_relation opens up to 50 at a time; a connection makes at most 120 calls a minute, 40 of them writes, and an account 240 and 80 across all its connections.

Sending a document to Filetable

MCP carries no files: what an assistant sends a tool, it writes, character by character. It reads the attachment you give it, but does not hold its bytes. Three paths spare it from copying out a PDF: ChatGPT hands Filetable the attachment itself, which Filetable downloads; an assistant that runs commands and reaches the internet sends the file to a single-use address; otherwise it gives you a link where you drop it, the path that works everywhere.

Examples, call by call

These five requests are the ones of a freelancer or a small business owner who keeps their own invoices, receipts, contracts and bank. The tables, columns and playbooks they name are the ones a new Filetable workspace is born with; in yours, the assistant reads the names you gave them.

What is left to pay

The assistant reads the table's shape first, because a filter names columns that exist. The payment status is not a box someone ticks: Filetable works it out from the invoice and what settles it, a payment date or a bank line. The total comes from aggregate_table: Filetable adds up every row that passes the filter, rather than the model adding up a page of results.

A receipt from a phone photo

A photo or a scan is read by OCR, like a PDF: Filetable takes its text, the QR-bill's code included, then the values the table asks for, each linked to where it was read on the document. The assistant waits for the reading to finish with get_processing_status before it answers. In Claude, the same receipt arrives through the upload link the assistant gives you (see above).

A bank line matched to its invoice

When a statement comes in, as camt.053 or CSV, Filetable's matching rules link a line to the documents they find on their own. What is left reads Unmatched, and that is where the assistant comes in: it passes the playbook's step as you would in its panel, required fields are checked, and both statuses follow the data. For an invoice paid before the statement arrives, Record the payment on the invoice is enough: one step, the payment date.

The contracts to cancel before year end

Filetable read each contract when it arrived: the end of the term, the notice period and the amount are in its row, and the notice deadline follows from them. The values are the ones you see in the app; if someone corrected them, the assistant reads the correction.

A table for a new kind of document

The table gets a description, which classification reads to send it the right documents, and each column its own, which extraction reads to know what to look for. The columns are created together or not at all. Relations between tables, computed columns and virtual folders are set up in the app.

More requests that work

• "Which customers are paying me late, and by how many days?"
• "How much VAT did I pay on my purchases this quarter?"
• "Which bank lines are not matched to anything, and which invoice could each one pay?"
• "What needs my attention right now?"
• "I deleted the wrong folder this morning, bring it back."

Security and control

• Your rights, no more. Every call goes through the same checks as a click in the app, workspace by workspace, table by table.
• No secret to keep. Claude and ChatGPT never receive your password; Filetable keeps only the hash of the keys it issues.
• You stay in control. In Settings → MCP, each connection shows when it was last used; pause it, change its workspaces or revoke it.
• Careful with deletions. Deleted files and folders go to the trash and can be restored for 30 days. A deleted row does not come back, and the assistant deletes only rows it named one by one; a deleted table is gone for good, with its rows. Read only, none of this is possible.
• A capped rate. Past 120 calls or 40 writes a minute for a connection, or 240 and 80 for an account, Filetable refuses the next calls and tells the assistant when to try again.
• A hidden row stays hidden. A row you cannot see in the app (its document sits in a folder closed to you) does not exist for the assistant, to read or to write.

What the assistant cannot do

Some things stay in the app, on purpose, because a connection must never be able to widen its own access or commit the account:

• Your account: password, email address, two-step verification.
• Members, invitations, roles and permissions.
• Billing, plans and credit purchases.
• Creating or widening a connection, or emptying the trash for good.
• Deleting every row of a filter in one go: each row has to be named.

If something goes wrong