Your AI on the go (remote connector)

Your books live on your computer — but with cloud sync the cloud keeps a synced copy, and the remote connector lets an AI app reach that copy when you’re away from your machine: ask for figures from your phone, or capture an expense as a draft the moment it happens.

The connector speaks MCP (Model Context Protocol), the same standard your local setup uses. It works with Claude (mobile, web, desktop, and Claude Code) and ChatGPT.

Other MCP-capable apps aren’t supported yet. An AI app has to identify itself in the way the current MCP specification asks for, and IrisBooks only accepts apps whose identity it recognises — so if your app can’t complete the sign-in, that is usually why, rather than anything wrong with your account.

Set it up

1. In your AI app, add a custom connector with this URL:

   https://irisbooks.io/mcp
   

   (In Claude: Settings → Connectors → Add custom connector.)

2. Your browser opens the IrisBooks sign-in and asks what to allow:

   • Bookkeeper (default) — the AI can read your books and create draft entries.
   • Viewer — read-only: reports and browsing, no drafting.

   If the app runs on your computer (Claude Code, for example), the page also warns that the approval goes to a program on this computer. Any program there could ask under that app’s name, so approve only if you just started the connection yourself.

3. Approve, and the connector is live in that AI app — including on your phone if the app syncs connectors across devices.

Your per-book roles still apply on top: even a Bookkeeper connection can’t draft into a book where you are only a Reviewer.

What you can do from anywhere

• Ask for figures — trial balance, P&L, balance sheet, monthly asset movements (cashflow is the tool’s legacy report name). This totals increases and decreases across all asset accounts, including receivables and fixed assets; transfers can appear on both sides. It is not a cash-flow statement. Numbers come from the cloud copy and include posted entries only; the AI is told when they were last synced from your machine and will say so.
• Inspect ledger and grouped totals — get_report(report="ledger", account=...) reads the account ledger; report="sum", by="tax.category,tax.rate" groups posted lines, including concern fields. Supply from/to, or year for sum.
• Read Net Worth and prices — get_networth uses your included books (or one book_id); get_holdings reads a book’s unit quantities; list_prices reads one page of your recorded prices. It accepts unit, currency, year, limit and cursor; follow nextCursor with the same filters. The first two accept as_of. These reads use synced data. Price recording and price sync stay in local MCP/CLI; the remote connector exposes no price write tool.
• Browse and search entries — by date, payee, amount, status.
• Check the fiscal years — which years exist, which are closed or reopened, and how many drafts still block a year’s close. (Closing and reopening themselves stay in the web app’s Year-end screen and the CLI.)
• Capture expenses as drafts — “I just paid ¥3,400 for a taxi” becomes a draft-status draft with proper accounts and balanced lines. It appears in your local book folder at your next iris sync, ready for review.
• Send in the receipt — the AI can mint a one-time upload link (works once, expires in 15 minutes, images/PDF up to 10 MB). The file is scanned, lands in raw/uploads/, and attaches itself to the draft. The upload page waits for that to finish and shows “the document is in your book” — that message, not the upload itself, is your receipt’s proof of arrival (it usually takes under a minute). If scanning is slow the page says so honestly; you can close it and reopen the same link later to check, and if the file is rejected (for example by the malware scan) the page tells you instead of pretending it worked. If the document is already in your email, the AI hands you the book’s receiving address instead.

Cloud cross-book reports share one automatic generation per user per UTC day with the web app. Later requests reuse the same inputs, even with different dates or books. Responses show snapshot.generatedAt and snapshot.nextRefreshAt; changes to journals or prices appear at the next generation. To include synced changes sooner, sign in to the web app and choose Run aggregation on Dashboard or Net Worth. Remote MCP then reads that updated snapshot; it cannot trigger aggregation itself. Reporting currency and inclusion preferences can re-render stored inputs immediately. Current access is always checked. A failed automatic generation consumes the daily slot, but the user can explicitly retry from the web app.

What it deliberately can’t do

The connector drafts, it never posts. It cannot:

• post or approve entries,
• edit or delete anything that’s posted or closed,
• draft into a sealed fiscal year — sealed-year journals cannot be written until the owner reopens it (its entries show as closed when browsing),
• seal or reopen a fiscal year.

Review stays where it belongs — at your computer or in the web app, with the full validate → post loop. Every remote-drafted entry is labeled in the audit history, so you (and your accountant) can always see which entries arrived from a phone.

At the computer or on the go?

You don’t have to choose — set up both. An AI app that can see both (e.g. Claude Desktop) knows to prefer the local one when it’s available.

Disconnecting

Two ways; either works:

• In your AI app — remove the connector in the app’s settings; the app tells IrisBooks to revoke its access.
• In the web app — your Settings (avatar menu) → Connected apps lists every AI app you have connected, with its access level and when it was last used. Disconnect revokes its access server-side — useful when you no longer have the device or app at hand.

Revocation takes effect within a few minutes. A collaborator’s underlying book role is separate — owners can change or revoke it any time in the book’s Settings → Grants.