HoneyBee
Processing
Read-only by default. Writes only as drafts a person approves.
The HoneyBee connector is a remote Model Context Protocol server that lets Claude — or any MCP client — ask questions about your real HoneyBee ERP account: cash, invoices, stock, projects, approvals, payroll self-service. You connect it once, with your own HoneyBee login, and it answers from your own database with your own permissions.
Anyone who already runs their company on HoneyBee ERP — EPC and engineering contractors, solar and energy operators, trading and services SMEs in Bangladesh, Singapore and Germany. It is meant for the people who ask the business questions: owners, finance and project leads, and the staff who just want their own payslip or leave balance without opening a screen.
You need an existing HoneyBee account. The connector adds no data of its own and creates no second copy of your records — it is a window onto the system you already have.
Authentication is OAuth 2.1 with PKCE, with dynamic client registration, so there is no key to copy and no secret to store. You sign in to HoneyBee in your own browser, see exactly what the connection is asking for, and approve it there.
In HoneyBee, open Settings → AI Connection. The page shows the exact server
URL for your company. It is your own HoneyBee address with /mcp on the end
— for example https://your-company.example/mcp.
If you have several HoneyBee companies, /my/mcp on the
central address is a second, cross-company server that answers about all of them at
once. It is read-only with no setting that changes that.
In Claude, go to Settings → Connectors → Add custom connector and paste that URL. Claude registers itself with the server automatically; you are not asked for a client id or a secret.
Your browser opens a HoneyBee sign-in page. After you sign in, a consent screen states in plain words what the connection will be able to do — and what it will never be able to do. It reflects your company's actual settings: if drafting is switched off for your company, the screen does not claim drafting.
Try “what is my cash position”, “which invoices are overdue”, or “what is waiting for my approval”. Answers come from your live ledger, and when HoneyBee cannot stand behind a figure it says so instead of returning a number.
Every connected client is listed on the same Settings → AI Connection page with the person who authorised it and when. Revoking one there stops it immediately — the next call it makes is refused. Revoking does not require support, and it does not affect any other connection.
Most business connectors are shaped by what the API happened to expose. This one is shaped by what we are willing to let software do to a company's money. The boundaries below are enforced in code, on the server, at every request — not by prompting the model nicely.
Out of the box the connector exposes 39 tools that can only read. Whether a tool can execute is decided by the risk class it declares in the tool registry, not by a list of names somebody maintains, so a new writing tool cannot appear on the read surface by being forgotten.
A company can switch on a second tier of 31 write tools. Calling one of them creates nothing: it returns a preview of what would change plus a signed, single-use draft reference. The record does not exist until a person approves it.
There is no code path on which a payment executes in the same request that asked for it. A money operation is always drafted first, and the draft must state the amount, the currency, the paying account, the receiving party, what it settles and the resulting balance effect — six facts, all of them, or HoneyBee refuses to produce the draft at all rather than show half a transaction.
Committing it is a separate call that must name that specific draft, quote a proof string printed on it, and restate the amount and currency exactly. A mistyped figure is refused. “Approve the last thing” is impossible by construction. If anything about the money changed since the human read it, the confirm is refused rather than adjusted.
And only a real person can do it: an authenticated human in an OAuth session who consented to a scope that says so in plain words, on a company that switched money confirmation on, with a kill switch un-pulled. A machine credential never sees these tools at all — they are absent from its tool list, not merely refused.
Five operations are unreachable through the connector at every setting: creating a supplier payment, posting a journal voucher, posting a contra voucher, approving a document and bulk-approving documents. A person may never approve, through this channel, work that they themselves raised — that control is what makes everything else safe, and there is no flag anywhere that lifts it.
The connector acts as the person who authorised it. Every call runs through the same permission gate as the ERP screens, and the company's own database is the only thing it can reach. A colleague's payslip is refused for the same reason it is refused in the browser.
Every tool carries a plain-language title and a machine-readable hint saying whether it only reads or can write, so your client can tell before it calls. The read tools below are the default surface — what a brand-new connection sees.
Drafting (31 tools, off by default). Quotations, sales and purchase orders, goods receipts, expenses, leads, projects, tasks, employees, products, customers and suppliers, attendance and follow-ups. Each returns a preview and a draft reference; none of them writes.
Approving from the chat (3 tools, OAuth humans only). List the drafts waiting for a decision, approve one, or reject one. A machine credential cannot see these tools.
Money confirmation (off by default, per company). Supplier payments and cash/bank transfers, drafted and then confirmed under the rules in section 3 above.
All my companies (5 tools). A separate read-only server that lists your companies, aggregates pending tasks and payables across them, and drills into any one of them. A company that does not answer is named as missing — it is never quietly counted as zero.
The connector reads your workspace data to answer your questions, under the same terms as the rest of the platform. What HoneyBee collects, how it is used and stored, who it is shared with, how long it is kept and who to contact are set out in full in our privacy policy.
Support: support@ourhoneybee.eu · Privacy: privacy@ourhoneybee.eu · Or use the contact form.
New to HoneyBee? Start with HoneyBee ERP.