World-first · MCP-native accounting

Connect your Claude. Let it do your books.

Bring the Claude / Cursor / Codex subscription you already pay for. It reads your bank feed, proposes reconciliations, categorizes transactions, and verifies your P&L — you approve in batch. No new AI subscription, no walled-garden chatbot, no inference fees on your card.

Start free → How it works

claude desktop · mcp

# Your prompt to your own Claude:
> reconcile april banking for company 19

# Claude calls TP MCP tools:
tp_get_inbox     → 47 transactions
tp_get_coa       → 14 chart accounts
tp_get_rules     → 23 vendor rules
tp_propose_categorize × 38
tp_propose_split      × 6
tp_end_session   → queued for review

# You open TP, see the session card:
"Claude reconciled 44 of 47.
3 flagged for human review."

Why this is different

Other accounting platforms run their AI. We run yours.

QuickBooks runs Intuit Assist. Xero runs Just Ask. Both are vendor-controlled models you can't customize, can't trust, and can't take with you. We took the opposite bet.

🤖

Use the Claude you already pay for

No second AI subscription. Your TP plan stays at $29.99. Inference cost is between you and Anthropic / OpenAI — not us.

🛡️

Nothing posts without your approval

Every Claude proposal lands in a Pending Review inbox. You batch-approve the high-confidence ones, eyeball the rest. AI never silently writes to your books.

🧾

Audit trail per AI-touched entry

Every AI proposal carries the model, the session, the prompt excerpt, the confidence score, and the approver. CPAs can ask "which entries did Claude touch?" — and get an answer.

How it works

Four steps. One install. Zero ongoing AI bills from us.

Paste one URL

In Claude Desktop / Cursor: Settings → MCP → Add custom connector. Paste https://trailingpaper.com/mcp. No JSON editing, no token to copy.

Click Allow

Your browser opens to TP's consent screen with the exact scopes the client requested. Click Allow. OAuth 2.1 + PKCE handles the rest. Tokens auto-rotate every hour.

Tell Claude what to do

"Reconcile April banking." "Find anomalies in expenses." "Verify Q1 P&L matches the bank." Whatever you'd ask a junior bookkeeper.

Approve in batch

Open TP. The Pending Review inbox shows Claude's session: 44 proposals, ready to batch-approve or eyeball one-by-one.

What Claude can do

Read everything. Propose anything. Commit nothing without your approval.

Read tools

tp_list_companies — companies the user can act on
tp_get_inbox — transactions awaiting reconciliation
tp_get_chart_of_accounts — full COA
tp_get_category_rules — existing vendor → category rules
tp_get_pl — P&L for any date range
tp_list_proposals — what Claude has previously proposed

Propose tools (always need your approval)

tp_start_session / tp_end_session — group proposals
tp_propose_categorize — categorize a bank txn
tp_propose_match — match to an invoice
tp_propose_split — hybrid: part-invoice, part-bill, part-categorize
tp_propose_expense — new expense without bank-feed link
tp_propose_rule — vendor → category rule

Compatible clients

Works with every MCP-compatible AI agent.

Anthropic shipped the Model Context Protocol as an open standard. Anyone can implement a client. We don't lock you to one vendor.

🟧

Claude Desktop

First-class. The flagship integration.

⌨️

Claude Code

CLI — great for power users.

▶️

Cursor

If your bookkeeper happens to be a developer.

🔌

Any MCP client

Codex, OpenWebUI, Continue, custom — the protocol is open.

Example prompts

Real things you can ask your Claude.

"Reconcile April for Acme Brewing. Match anything that looks like a customer payment, categorize the rest using my existing rules, end the session with a summary."

Claude opens a session, pulls the inbox + COA + rules, proposes 44 entries, ends. You approve in TP.

"My P&L for Q1 looks $5k off from the bank. Find the discrepancy."

Claude pulls P&L + recent ledger entries, cross-references against unreconciled transactions, returns a finding.

"Show me every transaction over $500 in March that hasn't been categorized yet."

Pure read-only query. Claude returns a list with context.

"Anytime you see 'AWS' in a transaction, propose categorizing it as 5400 Cloud Infrastructure. Make a rule."

Claude calls tp_propose_rule. The rule shows up in your Pending Review for one-click approval.

Common questions

Bring-your-own-AI FAQ

Yes — this feature lets you use the AI subscription you already have. If you don't have one, you can still use TP normally; the AI integration just won't apply. We don't sell our own AI add-on, on principle.

No. Every proposal lands in your Pending Review inbox. Nothing changes the ledger until you click Approve. (Power-user mode lets you set per-rule auto-commit thresholds, but it's off by default.)

Whatever scopes the token you generate has. Read sees inbox + COA + rules + P&L for the company you scoped it to. Propose additionally lets the AI submit proposals. We don't expose customer SSNs or bank account numbers via MCP — only the data needed for bookkeeping.

You see it in Pending Review with the proposed entry, the AI's reasoning, and the confidence. You reject. The audit_log keeps a record — rejections are tracked too. The AI doesn't get to write the wrong thing twice.

Real MCP. Anthropic's open Model Context Protocol. We publish a stdio MCP server (open source — view source) you install with npx @trailingpaper/mcp-server or via the install snippet TP gives you. Same protocol Claude Desktop, Cursor, etc. use natively.

One click. The agent loses access immediately on the next request. We log every token usage so you can audit.