The Taokeh AI connector
It drafts every entry — you keep the final tap. Nothing posts without you.
This is the connector for your AI: link Claude, ChatGPT or another assistant to your live books and just ask. It’s the same AI accounting Taokeh you run every day, now reachable from your chat.
Read-only to start. It can’t post anything until you tap approve.
No new app to learn. You talk to your AI the way you already do — Taokeh is the connector behind it.
Upload the day’s receipts, a supplier bill or a bank statement — or just ask a question — in your AI chat, in English, BM or the rojak you actually speak.
Your AI reads your live books and turns the paper into draft entries — a bill, an invoice, suggested bank rows — each with the figure it computed for you to check.
Every draft sits inert in your review deck until you tap Approve, one card at a time. Nothing posts, receives stock or moves what you’re owed until you do.
The wall, once
The connector’s read tools run under a forced read-only transaction — and there is no post tool your AI can reach. It can draft and suggest; only your tap writes to the ledger.
Same connector, four different jobs. Every one ends the same way — with your tap.
Hand your AI the day’s paper before bed; wake to a tidy stack of drafts — tap approve over kopi.
9:40pm. He snaps the day’s receipts, the supplier bill and the handwritten sales note into the chat. By 7:20am there’s a neat stack of pending drafts in Taokeh, each with the photo beside the computed total, waiting for his thumb.
Where you tap approve
ApproveEvery item lands as an inert PENDING draft in your /go deck and the review strips. Nothing posts, receives stock or moves what you’re owed until you tap Approve — one card at a time.
Ask your books anything before the school run — they answer, honestly, and file nothing.
No paper, no drafts — just questions, answered live from her real numbers. She can even connect read-only, and the whole flow still works end to end.
Ask “how much SST this round?” and it answers only if you’re SST-registered — software behaviour, not tax advice (rules current as of July 2026; check LHDN/RMCD for your situation).
Where you tap approve
ApproveNothing to approve — this flow reads and files nothing. Your AI can’t chase, email or pay anyone: it hands you a short-lived invoice link and points you to the reminders toggle. You do the sending.
Your bookkeeper preps every entry; you just thumb-approve — nothing posts without you.
By day the bookkeeper does the doing — reads the supplier PDFs, the sales-note photos, the bank statement — and leaves a stack of drafts. By evening the towkay taps Approve or Reject, per card, on his phone. The doing is hers; the yes is his.
Where the human taps approve
ApproveOne shared workspace. Every draft and every bank row carries an “AI suggestion” badge until his tap. A one-tap approve link still forces his own Taokeh sign-in — the link is not a login, and it cannot post.
Before you lock up: did your bank and QR tie today, and what’s left to sign?
9:40pm. She uploads the DuitNow QR settlement and the bank statement page. Her AI stages every row, checks the arithmetic, and shows what’s mechanical versus what needs her eyes.
Where you tap approve
ApproveThe importer refuses unless opening + rows = closing (±RM0.05), and tells you the exact ringgit gap if not — it never invents a row to tie. Rows land SUGGESTED, never posted; you categorise and post every line. Honest scope: this ties the money that reaches the bank (bank + QR), not the cash tin.
Some of what people want from an “AI accountant” isn’t a feature anyone ships — it’s a sequence of honest reads your AI strings together. Here are the two asked for most, with the exact prompt and the tools behind it, so you can judge the mechanics before you connect anything.
There is no “management report” button in Taokeh, and nothing here writes you a report. What there is: a set of read tools that answer with figures from the ledger, and an AI that can call them in sequence and write the thing. The composition is the report.
Paste this into your AI chat
“Prepare my management report for July: P&L versus last month, the top profit drivers, who owes me and how overdue, my cash position and the 8-week forecast, and anything unusual in expenses.”
What your AI will call
income_statementTwice — it takes a from/to date range, so your AI asks for the month you are closing and again for the month before. Revenue, cost of goods sold, gross profit, expenses, net income.profit_driversThe bridge that explains the move: revenue delta, COGS delta, then the operating-expense accounts that shifted most — with disposal and FX one-offs pulled out so a lumpy asset sale does not read as a trend.ar_agingWho owes you and how overdue, by customer and bucket.cash_positionEvery bank account and cash pocket with its GL balance, plus the liquid total. Money the owner has advanced comes back separately as a liability — never folded into cash.cash_forecastThe forward eight weeks: open invoices at their due dates against open bills, payroll, PCB/EPF/SOCSO/EIS, loan installments and SST — week by week, with the lowest projected balance called out and every assumption listed.expenses_summaryWhat you have actually spent — today and month-to-date.Six calls, maybe eight if it goes back for a customer’s detail. Your AI picks the order, reads each answer, and then does the part no tool does: sets the two months beside each other, notices that gross margin slipped because unit cost moved rather than because you sold less, writes the headings, and hands you a document.
That document is your AI’s work, on your AI subscription, in your choice of model. Taokeh never runs a language model on the connector path — we answer with arithmetic off the double-entry ledger and nothing else. So the narrative is its judgement and can be argued with; the figures under the narrative are your books’, and every one of them is checkable against the same Reports pages you already open.
Ask it to do this on the 1st of each month and you have a management pack. Ask it to do it on a Tuesday afternoon because a banker called and you have the same pack, four hours later. Nothing was scheduled, built or configured — you asked a question in a longer sentence.
Anything the AI drafts along the way still waits for your tap. Reading your books to write you a report changes nothing in them.
Same shape, different job. No tool here returns a field called “anomaly” — the tools return figures, and the judgement about what looks wrong is your AI’s, made out loud, as a list for you to look at rather than a verdict.
Paste this into your AI chat
“Before I close July: scan for anomalies — expenses that look doubled, customers suddenly overdue, bank rows still unreviewed, stock oddities — and list what I should look at.”
What your AI will call
daily_briefThe wide sweep in one call: sales and expenses today and this month, cash, top debtors and overdue buckets, what is waiting for approval, tax position, low stock.search_expensesThe duplicate hunt. Posted paid expenses, searchable by reference, memo text, date range or amount range — two identical amounts a week apart on the same reference is exactly the shape a doubled bill makes.expenses_summaryThe month’s spend as a whole, so an odd total shows against the shape you know.ar_agingWho has slipped a bucket since last month.open_invoicesOne named customer’s unpaid invoices, oldest first — the follow-up call when a name in the aging looks wrong.bank_reconciliation_statusGL cash against the latest statement close, row counts by status, and the unreviewed backlog. A backlog sitting there on the 28th is itself the finding.low_stockProducts at or below their reorder point — a line that has gone quiet or a count that cannot be right.The honest boundary: MCP is a pull. Taokeh answers when your AI asks and never the other way round — we run no watcher, fire no timer, and send no alert of our own. So this check happens when it is asked for. If you want it standing, that is your AI client’s job, not ours: several of them support saved scheduled prompts, so “run my month-end scan on the 28th” lives in your assistant, where you can see it and switch it off. We would rather point at that than claim a sentry we do not run.
We run our own company’s books on Taokeh and do this same scan before every close — through this exact connector, with these exact tools. Figures pulled into the open answer for themselves.
Anything the AI drafts along the way still waits for your tap. A scan reads; only you post.
Every tool named here is a read tool registered on the live connector — the full catalogue is in the developer docs. See the full tool catalogue.
The honest small print, because the trust is the product.
A few minutes, once. Then you just talk.
In Claude or ChatGPT, add Taokeh as a custom connector and sign in — a one-click grant, no token to copy. Manual clients (Claude Code, Cursor) use an access token instead.
The first time, you approve read access on the Taokeh consent screen. That grant is yours, and you can revoke it anytime.
Each connection is bound to a single set of books. Run more than one company? Connect each, and switch by hand — there’s no login that spans them.
The connector is RM29/mo — included in your 14-day free trial, so you can feel it before you pay.
Free for 14 days — read your books and draft, no card needed. RM29/mo after. See it first, or wire it up in a few minutes.
Read-only to start. It can’t post anything until you tap approve.