Guide · Malaysia and Singapore

    Connect Claude or ChatGPT to AutoCount and SQL Account

    Your accounts clerk spends Monday morning pulling an aged-debtors report out of AutoCount, pasting it into Excel and emailing it to you. You already pay for Claude, powered by Anthropic, or for ChatGPT. It writes a decent email, but it cannot see a single figure in your books.

    The missing piece is an MCP server, the plug that lets an AI assistant read another system. As of 14 November 2026, we found no MCP server published by either AutoCount or SQL Account, so for now the connection has to be built. Your choices:

    An office accounting server linked through a secure gateway to an AI assistant, with invoice, stock and sales panels

    Your choices

    • A few reports a month: keep exporting to Excel and attach the file. No build cost, though check what your AI plan does with files you upload.
    • The same questions about live figures every week: build a read-only MCP server. The AI answers from today's data, and nobody exports anything.
    • You want the AI to draft invoices or purchase orders: add draft tools behind an approval step, once the read-only version is trusted. Your team stops typing them from scratch, and a person still approves each one.

    Part 1 walks through the read-only route. Part 2 is for IT managers.

    Product facts checked against Anthropic, OpenAI and Google documentation on 14 November 2026.

    Part 1: Seven steps to a read-only AutoCount or SQL Account connection

    What you can ask once it is connected

    Questions like these:

    • Which customers owe us the most, and for how long?
    • What is our stock balance for this item in each warehouse?
    • Which supplier invoices fall due this week?
    • How do this month's sales compare with the same month last year?

    Each one only reads. The AI looks the answer up and changes nothing.

    Step 1: find out where your accounting data lives

    Start with your IT support or software dealer, and ask two things. Where does the database sit? Can a separate program read it safely?

    What the AI can reach depends on how your AutoCount is installed and what access it allows; in our experience the same holds for SQL Account. Does yours sit on a server in your office, or is it hosted by your dealer?

    If your dealer says no program can read the data safely, stop. An MCP server needs a way in: an API (a published way for other software to read the data) or a database it can reach. If your set-up has neither, a good builder tells you before you pay for anything.

    Step 2: choose the first question and keep it read-only

    Pick the report your team pulls most often. Read-only means the AI cannot post or change anything in your books.

    Start there even if you want more later. Widening access that works is easy; undoing a wrong entry in your books is not.

    Step 3: choose the AI your team will use

    As of 14 November 2026, Claude supports this on every plan; ChatGPT supports it on the web, though we could not confirm which plans; Gemini mostly does not for businesses here.

    Claude. We recommend Claude Team or Enterprise for finance teams. Your Claude administrator (Claude calls the role Owner) adds the connection once for the company, and each person signs in with their own account. Claude calls this a custom connector. As of 14 November 2026, Free, Pro, Max, Ultra, Team and Enterprise plans can all add one; Free is limited to one.

    ChatGPT. OpenAI's developer docs describe adding a custom MCP server in ChatGPT on the web. Account and workspace policies decide who can add one. We could not confirm which ChatGPT plans include it, so check before you build.

    Gemini. Not yet for most businesses here. The consumer Gemini app's custom MCP feature is limited to personal accounts in the US; Gemini Enterprise Business edition supports it in pre-release.

    A server hosted online (a remote server, rather than one on a single office PC) can usually serve both Claude and ChatGPT, so this choice does not lock you in.

    Step 4: decide who sees what

    Your sales manager should see sales figures, not payroll, even when asking the same question as your finance manager.

    We recommend per-user sign-in, where each person signs in with their own account and the server passes that identity through to the accounting data. If your software cannot support that, the server itself must limit which records each tool returns.

    Step 5: build the server and test it on real month-end questions

    This is the step a developer does: your IT team, your software dealer, or a firm like ours. We recommend designing the server with a small set of named tools, such as "outstanding invoices by customer", and nothing broader. In MCP terms these are tools the server offers the AI.

    Then test with the awkward cases your team really meets, such as a customer with two accounts or a credit note. Check every answer against the figure in AutoCount or SQL Account. Fix what it gets wrong before anyone relies on it.

    Step 6: add drafting, behind an approval step

    Once the answers are trusted, add tools that prepare drafts, such as a quotation or a payment reminder. A person checks each one before it goes anywhere.

    As of 14 November 2026, ChatGPT asks you to confirm before write actions, meaning anything that changes data. Build your own approval step as well, so the rule does not depend on one app's setting.

    Step 7: roll out to one team and keep a log

    Give access to the team that asked for it, not the whole company. Log every question and tool call for the first month.

    Keeping the AI's permissions in your own systems helps make its access reviewable under the Personal Data Protection Act (PDPA) in Malaysia and in Singapore, where the PDPC oversees it; the log helps you answer questions about it later. Treat this as a starting point, and take legal advice for your own case.

    How much it costs

    MCP is an open-source standard, so there is no protocol licence to buy. The cost is building and running the server.

    As a planning example from our own estimates (14 November 2026): one AutoCount or SQL Account connection, read-only, 11 to 50 users, hosted in your own cloud, with standard security and one AI client, comes to about 18 to 30 person-days, or 5 to 8 weeks. With per-user sign-in, as Step 4 recommends, it is about 21 to 35 person-days, or 6 to 9 weeks. We quote a price in ringgit or Singapore dollars only after checking what your set-up allows. Each extra, such as drafting or a second system, adds to it. Our AI connection planner helps you scope your own case, with a brief you can pass to your finance director. The AI seats are separate; our Claude cost calculator covers those.

    Plan your AutoCount or SQL Account connection

    Six questions turn into a brief: the first use case, a safe starting scope and what to prepare.

    Open the AI connection planner

    Part 2: For IT managers and developers

    Where the MCP server runs

    Claude reaches custom connectors from Anthropic's cloud, not from the user's PC. A server sitting next to an office-based accounting database is therefore not reachable from Claude on the web unless you plan for it:

    • A local server in Claude Desktop (a desktop extension), running on the user's own computer. Simple, but per machine, and Owners can restrict which extensions are allowed.
    • A remote server in your own cloud that reaches the office database over a secure connection you control. ChatGPT also expects a remote server.
    • MCP tunnels, an outbound-only connection from your network, in research preview for Claude Enterprise by request as of 14 November 2026.

    Transport and sign-in

    In the current MCP specification (2026-07-28), remote servers use Streamable HTTP. Authorisation is optional; when used, it is based on OAuth 2.1. Never put tokens in the URL query string.

    Mark read-only tools as read-only. As of 14 November 2026, ChatGPT treats tools without a read-only hint as writes and asks for confirmation by default.

    Do not point the AI at the raw database

    The official MCP reference servers, including the archived PostgreSQL one, are examples, not production code. We recommend against a generic "run any SQL" tool: it can hand the model everything its database login can read. Write narrow tools that return only the fields each question needs.

    FAQ

    Questions finance managers ask

    Can ChatGPT connect to AutoCount?

    Usually, yes, if your AutoCount set-up allows a program to read it. We found no MCP server published by AutoCount as of 14 November 2026, so one is built for your set-up and added in ChatGPT on the web. Check that your ChatGPT workspace allows custom MCP servers.

    Can Claude connect to SQL Account?

    Usually, yes. SQL Account is one of the systems we connect Claude to, through an MCP server added as a custom connector. As with AutoCount, it depends on what access your set-up allows.

    Is there an official AutoCount or SQL Account MCP connector?

    We found none published by either vendor as of 14 November 2026. If one appears, it may be cheaper to use than a custom build, and we would say so.

    Can the AI post entries into our books?

    It can be built to, but we do not recommend starting there. Read-only first, then drafts a person approves.

    Is our accounting data safe with an AI?

    It can be, if access is designed rather than assumed. The MCP specification says the protocol cannot enforce consent or access control, and that builders should. Read-only defaults and per-user sign-in keep permissions in your own systems, which helps make the access reviewable; a log helps too.

    Does this work for Singapore companies?

    Yes. In our experience the steps are the same for Singapore companies on AutoCount or SQL Account. On QuickBooks Online, see our read-only QuickBooks connector instead.

    Free consultation

    Want Claude or ChatGPT to read your AutoCount or SQL Account?

    Tell us how your accounting software is installed and the report your team pulls most. We will tell you what is reachable before you spend anything.

    Powered by EzyChat