OpenAI launched ChatGPT for Accounting on 18 November 2026 with eight named partner apps. Four of them can write to the books from a chat window, one writes as far as the signed-in user's NetSuite role allows, and three only read. Accounting systems outside those eight are left with computer use, where ChatGPT operates the app's screen the way a person would, unless a workspace admin adds a connection of their own.
A plugin is a typed, permissioned connection to the ledger. Computer use is a model reading pixels and clicking buttons in software that was never built for it. For product teams at accounting platforms and for anyone building agents on top of their customers' books, the useful question is which of those two doors a given ledger has today.
The launch: one plugin and eight apps
The launch has two parts. The first is an OpenAI-built Accounting plugin, a set of instructions for month-end work such as reconciliation and invoice entry. The second is the partner apps that give it access to records. According to OpenAI's help article, the plugin runs in the ChatGPT desktop app. Members of a Business workspace can install it themselves, while in an Enterprise workspace an admin has to approve it first. The solutions page states the fallback plainly: computer use "lets ChatGPT help in apps that aren't available through a plugin."
ChatGPT was not first here. On 28 July Intuit moved its QuickBooks connector in Claude and its plugin in ChatGPT beyond read-only access, to creating and sending invoices. On the consumer side, Finances in ChatGPT lets US users connect their bank accounts through Plaid, and it cannot move money. Plaid expanded the same kind of connection for Perplexity in April. Finance data is arriving in AI chat from several directions at once, and the accounting launch is the one that touches a company's general ledger.
Read access and write access are different products
OpenAI's own help article warns that "a connection that lets you read transactions may have different permissions for entering records." The partner listings bear that out.
FreshBooks says its connection "cannot change anything in a FreshBooks account", and its support pages describe write scopes as coming soon. Xero's ChatGPT listing says "read-only access". When Xero's announcement talks about moving "from insight to action back in Xero", the action happens in Xero's own interface. Xero keeps the ChatGPT plugin narrower than its own remote MCP server, which has been generally available since 13 October and includes write tools. BILL's MCP server is in beta and read-only, and its docs say paying bills is not available yet.
The partners that do write have put controls around it. QuickBooks keeps the user as the final approver before an invoice or estimate goes out. Campfire's launch post says approval requirements and closed-period controls apply to any change, and every call goes into an audit history. The Rillet API goes further when an approval workflow is on: a new invoice comes back with the status PENDING and only becomes a record once someone approves it.
The most practical line in OpenAI's documentation is about duplicates: "If an entry's status is unclear, check the accounting app before trying again to avoid creating a duplicate." A person reading a screen notices when an invoice already exists. An agent that retries a timed-out request does not, unless the integration underneath gives it a way to tell. Anyone writing to a ledger from an agent should treat retries and idempotency as part of the design, because the model will not catch a double-posted bill on its own.
The ChatGPT for Accounting tracker
We checked 21 accounting platforms against the AI assistant directories and the vendors' own developer docs on 20 November 2026. "Not found" means we found no listing on that date, not that one can never exist. We will keep this table current as vendors ship.
| Platform | In ChatGPT for Accounting | ChatGPT access | Claude directory | Vendor MCP server |
|---|---|---|---|---|
| QuickBooks Online | Yes | Read and write, US only | Read and write | Open-source server on GitHub (read and write) |
| Xero | Yes | Read only, US for this launch | Read only | Remote MCP, generally available, read and write |
| Oracle NetSuite | Yes | Read and write within the user's role | Read and write | AI Connector Service, role-based |
| NetSuite QuickStart | Yes | Read and write | Read and write | Included |
| BILL | Yes | Read only, US organizations | Read only | Beta, read only |
| FreshBooks | Yes | Read only, US, Canada and UK | Read only | Hosted, read only |
| Rillet | Yes | Read and write | Read and write | Read and write |
| Campfire | Yes | Read and write, with approvals | Listed | Read and write (documented tools read only) |
| MYOB | No | Own app, read only, Australia and New Zealand | Beta, read only | Beta, read only |
| Zoho Books | No | Not found | Read and write | Read and write |
| Microsoft Dynamics 365 Business Central | No | Not found | Not found | Built in, read by default, writes if an admin allows |
| Odoo | No | Not found | Not found | Built in (Odoo Online), read by default |
| Sage Intacct | No | Not found | Not found | AI Gateway MCP server, read only |
| Visma Business NXT and Visma Net | No | Not found | Not found | Read and write |
| Pennylane | No | Not found | Read only | Read only, write in progress |
| Exact Online | No | Not found | Not found | Not found |
| DATEV | No | Not found | Not found | Not found |
| Fortnox | No | Not found | Not found | Not found |
| Visma e-conomic | No | Not found | Not found | Not found |
| Sage Business Cloud Accounting | No | Not found | Not found | Not found |
| Wave | No | Not found | Not found | Not found |
The table splits into three groups. Eight apps sit inside the launch. A middle group sits outside the launch but publishes an MCP server, which a ChatGPT workspace admin can add as a custom app and which other agents can call directly. Only MYOB in that group has its own ChatGPT app, and it reads only. The last group has neither, and it is mostly European, with DATEV and Exact Online both in it.
The US-first problem
Half of the launch partners limit their ChatGPT connection by country. Intuit's help page says the QuickBooks connection is available to QuickBooks Online customers in the US. Xero describes its ChatGPT for Accounting integration as currently available in the US. BILL requires a US-based organization. FreshBooks adds Canada and the UK. We found no statement from OpenAI about when, or whether, the Accounting plugin will reach European customers.
For a bookkeeper in Rotterdam or an accounting firm in Munich, that leaves computer use or a custom connection. Europe is also where structured e-invoicing is becoming mandatory country by country, and Belgium made it compulsory between Belgian businesses on 1 January 2026. Structured, regulated documents are the work a model clicking through a screen suits least.
Computer use is the side door
The Financial Data Exchange published an essay on front doors and side doors earlier this month. Its argument is about consumer banking: agents should reach financial data through purpose-built APIs, not by logging in with a customer's password and scraping what they find. FDX is now working on agent identity and traceability, and on the tooling interfaces where protocols like MCP sit between an agent and an API.
The same reasoning carries over to business ledgers. An API returns an invoice with typed fields and a status. A screen returns whatever the layout shows, and the layout changes with every redesign and language setting. Computer use is a reasonable bridge for a long tail of niche tools. For the general ledger a company files its taxes from, it is a weak foundation, and OpenAI's own setup keeps a person approving every entry.
Becoming the ninth app
OpenAI built its app platform on MCP, and its submission guide and plugin guidelines are specific. The MCP server has to run at a stable public HTTPS endpoint on a verified domain. Each plugin connects one MCP server. Reviewers need a demo account with sample data, and plugins that require a new sign-up or two-factor login to test will be rejected. Tools must do what their names say, with no hidden side effects. Approval does not guarantee a directory spot either: an OpenAI support reply on its developer forum says approval and placement are separate and that placement is driven by usage signals.
For an accounting platform, the work splits into the protocol and the accounting. The protocol part is a remote MCP server with OAuth. The accounting part is what the launch partners have already shown matters: read scopes that are useful on their own, writes that return a preview or a pending state before they post, respect for closed periods, an audit trail of every call, and one sign-in that reaches several companies, which is the feature Xero highlights for accountants with a client portfolio.
If you are building on your customers' books
Most teams reading this are not accounting vendors. They build products like spend management or lending software, and their customers run on a mix of the ledgers in the table. Waiting for every platform to ship its own plugin means waiting on the last group, which has announced nothing. Building against each vendor's API or MCP server means a separate auth model per ledger, and a separate idea of what read-only means.
We built the Apideck Accounting API for that case: one API and one data model for the ledger, from invoices to journal entries, across the platforms in all three groups above, including several with no MCP server of their own. Some of those connectors are still in beta, and our connector list shows which. Your product calls the same endpoints whether the customer runs on QuickBooks or on Exact Online, and live reads versus a synced copy explains how the data reaches you.
For agents, the Apideck MCP server runs on the same Accounting API. Permissions are set per connection with read, write and destructive scopes, and an agent without the destructive scope cannot see the delete tools at all. Our QuickBooks MCP page and the NetSuite MCP post compare vendor and unified servers for those two ledgers.
If your product needs to read or write your customers' books, start with read access on the ledgers your customers use most through the Accounting API, and add write access once your retry and approval logic can handle the duplicate problem OpenAI describes.








