How it works

Your agent does the books. The ledger keeps it honest.

LedgerMCP has three parts: the AI assistant you already use, an MCP server that gives it the bookkeeping tools, and a double-entry ledger that enforces the accounting rules. Here is how they fit together, end to end.

Start free, no cardRead the tool reference

Your agent, any chat

Set up books for my LLC and import this statement.
Books created with a standard small-business chart. Imported 62 transactions, categorized 59, flagged 3 for you. Say the word and I’ll reconcile against the statement balance.

The three parts

Your agent

Claude, ChatGPT, Grok, or any MCP client. You talk to it in plain language: “import this statement,” “what did I spend on software?” It plans the work and calls the tools. You bring the assistant you already trust and pay for, or use the optional built-in assistant on Growth and Scale.

The MCP server

One URL, ledgermcp.com/mcp, exposes the whole job as 66 tools: 25 that read, 41 that write. Import, categorize, split, transfer, reconcile, tag, report. The same registry generates the docs, so what the agent sees never drifts from what the server does.

The ledger

Real double-entry books in a database that won’t let them break: every entry has to balance, nothing can be quietly edited or deleted, and the same thing can’t post twice. How, exactly →

Four ways to connect

1

Claude

In Claude: Settings → Connectors → +, then search for LedgerMCP and click Connect. It is in Claude’s connector directory, so there is no URL to paste. Sign in with an email and a 6-digit code and Claude has the full toolset. Works on claude.ai, Desktop, and mobile.
Full Claude walkthrough →
2

ChatGPT

In ChatGPT: Settings → Connectors → Advanced → developer mode, then Create and enter the URL. OAuth sign-in, no keys to paste. Available on Plus, Pro, Business, Enterprise, and Edu plans.
Full ChatGPT walkthrough →
3

Grok

In Grok: grok.com/connectors → New Connector → Custom, then paste https://ledgermcp.com/mcp as the server URL and sign in when prompted. Grok discovers the whole toolset from the server, same as every other client.
4

API key

For Claude Code, desktop agents, and scripts: mint a key in Settings → Agent access and send it as a bearer token. One key covers every book in your account.
MCP server details →
Connect your agent
{
  "mcpServers": {
    "ledgermcp": {
      "type": "http",
      "url": "https://ledgermcp.com/mcp",
      "headers": { "Authorization": "Bearer lmcp_YOUR_KEY" }
    }
  }
}

The config on the left is all an MCP client needs.

A month in the life of your books

  1. 1

    Statements arrive

    Bank feeds pull transactions in automatically (paid), or you hand your agent a CSV or PDF statement (free). Either way, rows land in a review queue. Nothing posts to the ledger by itself.

  2. 2

    Your agent does the work

    It categorizes against your chart of accounts, splits mixed charges, matches transfers so nothing double-counts, flags likely duplicates, and tags vendors as it goes. Anything ambiguous is left flagged for you.

  3. 3

    You review and close

    Skim the review queue in the web app or ask the agent what it was unsure about. Then close: reconcile against the statement balance, book depreciation, and lock the period so nothing can post into a closed month.

When something goes wrong

The agent miscategorized

Tell it (“that Costco charge was inventory, not meals”) and it reverses and re-posts. Or fix it yourself in the app with one click. Either way the correction is a linked reversal, and the suggestion engine learns the merchant for next time.

Something looks off

Open Settings → audit log. Every agent write is there: which key acted, which tool ran, what entry resulted. One click reverses any of it. There is no state where you can’t see exactly what happened.

You want a tighter leash

Issue a read-only key: the agent can answer questions and run reports but can’t write a thing. Or skip keys entirely and work only through the web app. The books are identical either way.

Common questions

Do I have to use an AI agent?

No. The web app is a complete bookkeeping tool on its own: transaction register with inline categorization, journal, reconciliation, receipts, tags, and every report. The agent just automates the same books.

What does the AI actually control?

Exactly what the tools expose. It cannot edit or delete postings, and destructive actions stay human-only in the web app.

Can I start mid-year with months of backlog?

Yes. Catch-up is one of the best uses. Hand your agent the statements month by month; bulk import and bulk categorize were built for migration, and one bad row never blocks the rest. Most backlogs clear in an afternoon.

What happens to my books if I stop paying?

Nothing happens to the books. Paid plans cover bank feeds and the built-in assistant; if a plan lapses those pause, but your books, reports, and agent access all keep working, and you can export everything to CSV at any time.

See it with your own statement

Hand your agent one month of transactions and read the audit log after.

Start free, no cardRead the docs