Connect Karchu to Claude and ChatGPT

What this is

Karchu turns bank statements into clean, categorized transactions. Most people then ask questions about that data: what did I spend on software last quarter, which subscriptions renewed twice, how much came in from one client, what is missing before I file taxes. Answering those questions used to mean exporting a CSV and pasting it into a chat window, which is slow, loses structure, and hands a full copy of your finances to whatever tool you pasted it into.

The Karchu MCP server removes that step. MCP, short for Model Context Protocol, is an open standard for letting an AI assistant call tools on a service. Karchu publishes a small set of read-only tools. You connect once, approve the connection with your Karchu login, and from then on your assistant can query your own data directly and answer in plain language, with references back to the transactions it used.

The connection is per person. It carries your identity, not a shared API key, so an assistant sees exactly what you would see if you signed in to Karchu yourself, and nothing else.

Server address

https://karchu.com/mcp

Transport is Streamable HTTP. Authentication is OAuth 2.1 with dynamic client registration, so most assistants can register themselves and simply send you to a Karchu sign in page. There is no key to copy and no token to paste.

Available tools

Seven read-only tools ship today. None of them can create, change or delete anything in your account.

Tool	What it does
`list_organizations`	Lists the workspaces you belong to and your role in each.
`list_transactions`	Recent transactions with an optional date range and text search.
`search_transactions`	Filtered search: date range, amount range, keywords, money in or out, category, review status, sorting and paging.
`get_transaction`	One transaction in full: category, the statement document it came from, linked receipts, and recent edits.
`financial_summary`	Money in, money out, net flow and the top spending categories for a period.
`list_documents`	Uploaded statements with processing status, bank, period and transaction count.
`list_categories`	The category taxonomy in your workspace, including your custom categories.

The search tool is the one most conversations end up using. It accepts a date range, an amount range, keywords matched against both the description and the merchant, a direction (money in or money out), a category by name, a review status, sorting and paging. That combination covers the majority of questions people ask about a statement without the assistant needing to pull down the whole ledger first.

Connect Claude

In Claude, open Settings and find Connectors (in some versions this sits under Integrations). Choose the option to add a custom connector, paste https://karchu.com/mcp as the server URL, and confirm. Claude registers itself, then opens a Karchu page in your browser.

Sign in with the account that holds your statements, read what the connector is asking for, and approve. Claude shows the connector as connected and lists the seven tools. Ask something like "using Karchu, what did I spend on software between January and March" and it will pick the right tool on its own.

Claude Desktop works the same way through its connector settings. If you are running a local configuration file instead, point an MCP remote entry at the same URL; the OAuth flow opens in your default browser.

Connect ChatGPT

In ChatGPT, open Settings, then Connectors, and add a custom or remote MCP connector. Paste https://karchu.com/mcp, continue, and complete the Karchu sign in and approval screen when it appears. Availability of custom connectors depends on your ChatGPT plan and whether your workspace administrator has enabled them, so if you do not see the option, that is where to look first.

Once connected, ask questions the same way. Naming Karchu in the prompt helps the model choose the connector on the first try: "check Karchu and list every transaction over 500 last month".

Connect Cursor, Windsurf and other clients

Any client that speaks MCP over Streamable HTTP can connect. In Cursor, add an MCP server entry with the URL above and approve the browser sign in when it opens. Editors and agent frameworks that expect a JSON configuration usually take a shape like this:

{
  "mcpServers": {
    "karchu": {
      "url": "https://karchu.com/mcp"
    }
  }
}

Clients that only support stdio can use a generic remote MCP bridge package and pass the same URL. Nothing about the server is client specific.

What the sign in screen means

The page you land on after clicking connect is served by Karchu, not by the assistant. It names the client asking for access and asks you to approve or deny. Approving issues that client a token tied to your account. Denying ends the flow and nothing is shared.

If you are not signed in when the connector sends you there, you will see the normal Karchu login first and come back to the approval screen afterwards. Sign in with Google works here too.

Example prompts

What did I spend on software subscriptions last quarter, broken down by merchant?
Show every transaction over 1,000 between April and June that is still unreviewed.
Which categories grew the most compared with the previous three months?
Find anything that looks like a duplicate charge from the same merchant in the same week.
Pull up transaction details for the largest expense in May, including which statement it came from.
Summarize money in and money out for the last financial year and list the top ten spending categories.
Which uploaded statements failed to process, and what was the error?

Because the assistant gets structured rows rather than a pasted table, the answers cite real transaction ids and you can open any of them in Karchu to check the source statement.

Security and access control

Three things keep this safe. First, the connection is authenticated as you, using a short lived token issued after you approve it, not a shared secret embedded in a config file. Second, every query runs through the same row level security rules that protect the app, enforced in the database rather than in application code, so a tool physically cannot read a workspace you do not belong to. Third, every tool is read-only. There is no tool that deletes a document, edits a transaction or changes a plan.

Statement files themselves are not streamed to the assistant. Tools return parsed metadata and transaction rows, so an assistant learns that a statement came from a given bank and covers a given period without receiving the PDF.

Logging and limits

Every MCP request is written to an activity log with the tool name, a summary of the filters used, the number of rows returned, how long it took, and whether it succeeded. You can read that log at Settings, then AI access. Workspace owners and managers see activity for the whole workspace, so you can tell who connected what and how often.

Fair use limits sit at 60 requests per minute and 1,000 per day per person. Going over returns a clear message the assistant can relay, and the attempt is recorded in the log so a misbehaving agent is easy to spot. The limits are generous for interactive use; if you are building something that needs more, get in touch through the contact page.

Revoking access

Remove the Karchu connector in your assistant to stop it using the connection. Changing your Karchu password or signing out everywhere invalidates the sessions behind it. Because access is per person, removing someone from a workspace also removes what their assistant can read from that workspace, immediately, on the next request.

How this differs from the public API

Karchu also publishes a public API. That API is anonymous and returns reference data: supported banks, accepted formats, the default category taxonomy, published benchmarks. It contains no personal data and needs no login. The MCP server is the opposite: it is private, requires your sign in, and returns only your own financial data. Use the API to build something on top of Karchu, use MCP to let an assistant work with your account.

Frequently asked questions

What is MCP?

MCP (Model Context Protocol) is an open standard that lets an AI assistant call tools on a service you control. Karchu ships an MCP server, so assistants like Claude and ChatGPT can read your statements and answer questions about them without you exporting anything.

Does connecting Karchu give an AI company my bank login?

No. Karchu never holds a bank login. You upload statements yourself, and the assistant reads the parsed results through your Karchu account after you approve the connection.

What can the assistant see?

Exactly what your Karchu account can see, no more. Access rules are enforced by the database on every request, so a member of one workspace cannot read another workspace's transactions through an assistant.

Can I see what the assistant asked for?

Yes. Settings, then AI access, lists every request: which tool ran, which filters were used, how many rows came back, how long it took, and whether it succeeded. Workspace owners and managers see the whole team's activity.

Is there a limit?

Fair use is 60 requests per minute and 1,000 per day per person. That is far more than a normal conversation needs, and it keeps a runaway agent from hammering your data.

How do I disconnect?

Remove the Karchu connector in your assistant, and revoke the session from your Karchu account. New requests stop immediately.

Next steps

If you have not uploaded a statement yet, start on the bank statement analyzer and run one file through. Once you can see categorized transactions in your dashboard, connect your assistant and ask it something you would normally answer with a spreadsheet. For background on how parsing and categorization work before the assistant ever sees a row, read how it works and the AI learning center.