We get one question more than any other, in demos and in support email: what actually happens when I type "record this bill" into Claude? Where does the AI end and DeepLedger begin?
This post is the full answer. We will walk through the three pieces of the system (the AI assistant, the MCP server, and the portal), then list every tool the server exposes, so you know exactly what the AI can and cannot do in your books.
The Architecture in One Paragraph
DeepLedger is not an AI. It is the layer between an AI and your books. The reasoning is done by an assistant you already use. Claude is the one we design around, though ChatGPT and any MCP-compatible client work the same way. That assistant connects to the DeepLedger MCP server, a hosted service at https://mcp.deepledger.ai/mcp that exposes QuickBooks Online as 26 structured, permissioned, logged tools. And sitting beside the chat is the DeepLedger portal: a web app built specifically for the human half of AI bookkeeping: working the shared task list, approving adjusting entries, signing off on the close. The AI and the portal share one database, so when the agent flags something in chat, it appears in your task list in real time.
Claude as the AI Accountant
We made an early decision not to train or host our own model. A frontier assistant like Claude already knows what an accrual is, why debits must equal credits, and what a reasonable chart of accounts looks like. What it lacks is access to your books, which is what MCP, the Model Context Protocol, was designed to solve.
So in practice, "using DeepLedger" mostly means talking to Claude in plain English:
- "Pull the bank feed and categorize what you can. Flag anything you're not sure about for my review."
- "Create a bill for the Staples invoice I just uploaded, due net 30."
- "Run the June close: reconcile, check AP and AR, propose accruals, and give me a trial balance."
- "Why is gross margin down this quarter versus last?"
You never call a tool by name. Claude reads the request, picks the tools, and chains them: looking up the vendor, checking for duplicates, then creating the bill. Two DeepLedger-specific things make it behave like a trained staff accountant rather than a generalist with API access:
Workflow guides. The server includes a getGuide tool with five built-in playbooks: transaction recording, month-end closing, financial analysis, audit preparation, and reconciliation. The first two are fully live today: each returns the exact sequence of steps, the safety checklist, and the common mistakes to avoid. When you ask Claude to close the books, its first move is pulling the closing guide, so every close follows the same procedure.
Per-client memory. The memory tool gives the agent durable knowledge about each company: patterns it has noticed, policies you have set, general notes. When you correct a categorization in the portal, that correction is written to memory as a policy, so the vendor that was miscategorized in March is categorized correctly in April, without you repeating yourself. You can read, edit, and delete everything the agent remembers on the portal's AI Memory page.
The MCP Server and All 26 Tools
Below is the complete tool inventory as registered on the server, grouped the way an accountant would group them.
Accounts Payable
| Tool | What it does | Access |
|---|---|---|
qb_bill | Create or update a bill (unpaid vendor liability), with line-level accounts or items, classes, and billable status | Write |
qb_bill_payment | Pay one or more bills from a bank or credit card account, including applying vendor credits | Write |
qb_expense | Record a purchase paid immediately by cash, check, or credit card | Write |
qb_credit | Create a customer credit memo (reduces AR) or a vendor credit (reduces AP) | Write |
Accounts Receivable and Sales
| Tool | What it does | Access |
|---|---|---|
qb_estimate | Create, update, close, or convert a quote/estimate to an invoice (non-posting until converted) | Write |
qb_invoice | Create or update an invoice, with terms, due dates, tax codes, and optional email delivery through QuickBooks | Write |
qb_receive_payment | Record a customer payment and apply it against open invoices and credit memos | Write |
qb_sales_receipt | Record an immediate sale (payment received at time of sale, no AR) | Write |
qb_refund_receipt | Record actual money returned to a customer | Write |
qb_send_email | Email a sales document (invoice, estimate, sales receipt, credit memo, refund receipt) to the customer and confirm delivery. The only tool that emails a customer, and always opt-in, never a side effect of recording | Write |
General Ledger and Banking
| Tool | What it does | Access |
|---|---|---|
qb_journal_entry | Create or update a journal entry; debits must equal credits or the tool rejects it | Write |
qb_deposit | Record a bank deposit with line-level accounts and customers | Write |
qb_transfer | Move money between bank accounts | Write |
qb_void_transaction | Void a transaction: zeroes the amounts but keeps the record for the audit trail. The only destructive tool on the server, and it cannot delete | Write |
Master Data, Queries, and Reports
| Tool | What it does | Access |
|---|---|---|
qb_company_profile | Read the company's identity and the QuickBooks settings that gate other tools: currency, sales tax, class and department tracking, book close date, report basis | Read-only |
qb_master_data | Retrieve, create, or update accounts, vendors, customers, items, classes, and tax rates | Write |
qb_fetch_transactions | Query transactions across every type: open-document checks, duplicate checks, account inference, report scans | Read-only |
qb_reports | Generate 19 report types: P&L, Balance Sheet, Cash Flow, Trial Balance, General Ledger, aged receivables and payables, sales by customer or product, budget vs. actuals, and more | Read-only |
qb_recurring_transaction | List, read, create, or delete recurring transaction templates | Write |
qb_attach_file | Attach a receipt, invoice, or document directly to a QuickBooks transaction | Write |
The DeepLedger Platform Tools
These six don't touch QuickBooks at all. They connect the agent to the DeepLedger workflow layer:
| Tool | What it does | Access |
|---|---|---|
bank_feed | Fetch bank transactions from your connected bank feed or CSV/Excel imports, with flags for items that already have a task | Read-only |
documents | Read uploaded documents from DeepLedger storage or attachments already in QuickBooks | Read-only |
tasks | The shared human/AI task list: escalate a bank transaction, document, or question to a human (written reasoning required), pull the tasks assigned to the agent, comment, hand off, and complete recorded items with the QuickBooks transaction ID | Write |
memory | Read and write durable per-client knowledge: patterns, policies, notes | Write |
closeRun | Drive the month-end close workflow: six steps, sixteen checks, proposed entries, financials, and a final score | Write |
getGuide | Retrieve one of five step-by-step workflow guides with safety checklists, plus a per-tool error recovery playbook | Read-only |
That is the whole list: 26 tools, 6 of them strictly read-only, and one void operation that zeroes amounts without deleting the record; in clients that support confirmation prompts, it checks with you before it runs. Eleven of the transaction tools also ship with an interactive card. When Claude records a bill payment, the result renders as a confirmation card showing the vendor, the bank account it was paid from, the check number, and the per-bill allocation, styled to match your chat client. These are MCP Apps, rendered inline in the conversation.
What's Enforced in Code vs. What's Convention
We wrote a whole post on what actually happens when you give an AI write access to your books, and the same standard applies here: some of DeepLedger's safety is enforced by the server, and some is built into the agent's instructions. The distinction matters.
Enforced in code: Access is granted through OAuth 2.1 with PKCE, scoped to the organizations you select at consent, and revocable at any time. QuickBooks tokens are encrypted with AES-256-GCM. QuickBooks permits one live app connection per company, so your team shares a single authorized connection, and DeepLedger's own audit log records which team member drove every action. There is no delete tool. Requests are rate-limited per organization. And the server logs tool usage as metadata only (tool name, duration, success), never your financial data.
Convention, enforced through instructions: The agent operates under an analyze-decide-execute framework: look up the entities involved, check for duplicates and open documents, and only write when the human has confirmed or the books clearly support it; otherwise, open a task for your reviewer. Every write tool's definition embeds its prerequisites (check for an existing open bill before creating one, for instance). Claude follows these reliably, but they are instructions to a model, not hard enforcement. The enforcement is the review workflow, which is what the portal is for.
The Portal: an Accounting Skin for AI Work
A chat window is a great place to delegate work and a poor place to review it. You cannot skim forty categorizations in a transcript, and "scroll up to find where it mentioned the accrual" is not a close procedure. So the DeepLedger portal exists as the review surface, an accounting-specific UI rendered from the same database the AI writes to:
Tasks are the heart of it. Everything the agent flags (uncertain bank transactions, documents it can't confidently classify, questions it needs answered) lands here as a task with the AI's proposed category, a confidence score, and its written reasoning. Every task shows whose turn it is: yours or the AI agent's. You can approve, reject, or edit the category before approving, singly or in bulk (approving hands the task back to the agent to record), and every action is written to the audit log. You can also create tasks of your own: a to-do for yourself or a job you hand straight to the AI agent. Corrections feed back into the agent's memory, so the list gets quieter over time.
The approve-then-record loop. Approved tasks don't post themselves: the agent picks them up through the tasks tool, records each one in QuickBooks, and completes the task with the resulting transaction ID. Human judgment gates the write; the AI does the data entry.
Month-End Close is a live view of the closeRun workflow. While Claude works through the six close steps, the page streams progress in real time. When it finishes, you get a scored checklist of sixteen checks, the exceptions it found, and its proposed adjusting entries (each flagged for materiality), which you approve or reject individually. The close ends with a sign-off: an owner or admin attests to the results, notes and all, and the period gets its closed certificate.
Everything else the human side needs: a bank feed page for connecting accounts and importing CSV or Excel files, a documents page for uploading the receipts and invoices the agent reads, the AI Memory page for inspecting and editing what the agent knows, and a dashboard summarizing review activity, cash flow, and vendor spend across every client company. Multi-company and multi-user are native: firms manage many client organizations under one login, with roles from owner down to bookkeeper, enforced at the database level.
Putting It Together: a Day in the Loop
Here is the whole system in one workflow. You upload a stack of receipts to the portal and tell Claude: "Process the bank feed and this week's documents." Claude pulls the transaction-recording guide, fetches the bank feed, reads the documents, and checks each item against QuickBooks history and its memory of your policies. The clear items (bills, expenses, deposits) it records directly, checking for duplicates first. The ambiguous ones it flags with reasoning. Ten minutes later you open your task list, bulk-approve the high-confidence flags, fix one category, and reject one duplicate. Next session, Claude picks up its approved tasks, records everything in QuickBooks, and completes each one. The correction you made is now a policy in memory. Both audit trails, DeepLedger's and QuickBooks' own, show exactly who did what.
That is how DeepLedger works: no proprietary model and no black box, just the assistant you already use, a documented set of 26 tools, and a human sign-off at the end.
Frequently Asked Questions
Is DeepLedger an AI? No. DeepLedger is the tool layer and review workflow; the reasoning comes from the assistant you connect: Claude, ChatGPT, or any MCP client.
How many tools are there, really? 26 registered tools on the MCP server: 20 for QuickBooks Online and 6 for the DeepLedger platform (bank feed, documents, tasks, memory, close, guides). Six are read-only. The list above is complete.
Can the AI delete things in my books? No tool deletes posted transactions. Void, which zeroes amounts but preserves the record, is the only destructive transaction operation. Recurring transaction templates are the sole exception: those are schedules, not history, and the AI can remove them.
Do I have to use the portal? The MCP server works from chat alone, but the portal is where tasks get worked, closes get signed off, and memory gets audited.
How do I get started?
Create a DeepLedger account, connect QuickBooks with Intuit's official OAuth, then paste https://mcp.deepledger.ai/mcp into Claude's connectors or ChatGPT's apps. The two-step setup guide walks through both clients.
If you want to see the loop for yourself, DeepLedger connects QuickBooks Online to Claude, ChatGPT, and any MCP-capable agent, with the full toolset above, per-client memory, human review built in, and a complete audit trail.
Try DeepLedger with your QuickBooks account or read the complete QuickBooks MCP server guide.