How do you connect Bexio to ChatGPT? Less simply than you might think, and on shifting ground: OpenAI announces the retirement of custom GPTs on 11 December 2026, in favour of plugins. The Bexio side does not change: a token opens all of the company’s accounts. Here is what exists today, what we could confirm and what you need to check in your own plan.
The short answer: do not build a custom GPT any more. The current route goes through an MCP server added to ChatGPT as a plugin, mainly open to Business, Enterprise and Edu plans. The Bexio token (60 days, all rights) then ends up on that server: choose it with care, start read-only and check the data settings of your plan.
How to connect Bexio to ChatGPT: what exists in October 2026
We found neither an official Bexio connector nor a Bexio app in the sources we consulted. For ChatGPT to talk to Bexio, two technical paths remain, one of which is closing.
| Custom GPT with Actions | MCP server added as a plugin | |
|---|---|---|
| Principle | ChatGPT calls Bexio’s API from an OpenAPI schema that you provide | ChatGPT connects to a server that exposes tools (list contacts, for example) |
| Status | Retirement announced for 11 December 2026 | OpenAI’s current route |
| Authentication | None, API key (Basic, Bearer, custom header) or OAuth | OAuth or none, according to the OpenAI pages we consulted |
| Where the Bexio token lives | In the configuration of the Action | On the MCP server, with its host |
| Plans | All are affected by the retirement | See below, to check in your workspace |
One caveat before going further: OpenAI’s help pages change quickly, and this table is dated 27 October 2026.
The Bexio side, in brief
Everything about Bexio is identical to what we verified for Claude. The detail is in our article How to connect Bexio to Claude. The essentials:
- the personal token is created at developer.bexio.com/pat, with your Bexio login;
- it lasts 60 days and carries all default scopes, hence full access to the company’s data: Bexio points to the OAuth Authorization Code Flow for restricting rights;
- it acts, according to Bexio’s API documentation, with the rights of the user who set up the connection;
- deleting it cuts access, but the full effect can take up to one hour;
- Bexio limits requests per minute and per company, and returns a 429 error beyond that;
- Bexio reserves this token for personal use;
- to share access with a third party, Bexio requires the OAuth flow, not this token;
- the role required to create it is not confirmed by the official documentation (one specialist guide mentions the super administrator).
One new point for ChatGPT: Bexio does not publish an OpenAPI schema. Its FAQ answers that no definition is available for now and that putting one online is planned. For a GPT with Actions, you would have had to write that schema yourself, endpoint by endpoint. One more obstacle on a route that is already doomed.
Route 1: the custom GPT with Actions is at end of life
A custom GPT can call an external API through Actions. OpenAI describes three authentication modes: none, API key (Basic, Bearer or custom header) and OAuth. The schema must be in OpenAPI format, in JSON or YAML. The Bexio token, which is a Bearer token, could therefore have been entered there as an API key.
But three facts change the picture:
- The retirement. Custom GPTs are scheduled to be retired on 11 December 2026, across all plans (11 February 2027 for some Enterprise workspaces with an approved deferral). For the affected Enterprise workspaces, creating new GPTs ended on 26 October 2026. An existing GPT stays usable until the retirement date, then becomes inaccessible.
- Actions do not transfer. When migrating to a plugin, instructions become a skill and knowledge files become reference files, but custom Actions are lost. OpenAI says they may require a custom MCP server and some technical setup.
- Token storage is not documented. OpenAI’s help page does not say where the API key is kept or who can see it. We therefore cannot claim anything about it. As a precaution, assume that anyone who can edit the GPT, and the workspace administrator, might be able to reach it.
If you already have a GPT linked to Bexio, do not share it, do not extend it, and prepare to delete it, along with the token. Note also that public GPTs with Actions must declare a privacy policy: a public Bexio GPT would be a very bad idea anyway.
Route 2: the MCP server added as a plugin
MCP (Model Context Protocol) is a standard that lets an AI call tools exposed by a server. In ChatGPT, you add it from the Plugins section, with “Add custom MCP server”. You enter the server address (“Server URL”) or a secure tunnel for a private server, configure authentication, accept a risk warning, then create it as a plugin.
What we could confirm in OpenAI’s pages:
- The plans. The help page describes full support for Business and Enterprise/Edu, on the web only (not on mobile). Pro can connect to servers with read and fetch permissions only. We found no official mention for the Free and Plus plans: check in your account.
- Who decides. On Business, only administrators and owners create and publish these plugins. On Enterprise/Edu, administrators control access by group and switch each action on or off. New actions are disabled by default.
- Authentication. The pages on custom servers describe OAuth or no authentication. The documentation for plugin developers also mentions an API key declaration sent as a header, but its scope is unclear. We therefore could not confirm that a Bexio personal token can be entered directly into ChatGPT.
- Workspace controls. Administrators can forbid adding plugins or custom servers. Your internal policy may therefore block you, and that is normal.
In practice, for Bexio, you generally need an intermediate server that holds the token and turns it into tools. Two consequences:
- The token is with the server’s host. Whether it is a provider, an automation tool vendor or yourself, that is where it sits. Yet Bexio reserves this token for personal use, without sharing it with a third party.
- A server with no authentication is open to anyone who knows its address. This is common-sense reasoning, not an OpenAI statement, but it is enough to rule out that setup for real accounts.
OpenAI says it itself: custom servers are not developed or verified by it, and an “OpenAI Verified” badge does not replace your own security and vendor review.
Privacy: what happens to what ChatGPT reads in Bexio
Everything a tool sends back to ChatGPT enters the conversation. As we read it, that data therefore follows the rules of the conversation, which depend on the plan.
Business, Enterprise, Edu and API plans. OpenAI writes that it does not use business data to train its models by default (feedback you choose to send is an exception). Retention is set by the administrator (Business) or by the customer (Enterprise, Edu); deleted conversations are erased within 30 days, unless the law requires otherwise. For the API, inputs and outputs are kept for up to 30 days, with zero retention available for some eligible access. A data processing addendum (DPA) is offered for Business, Enterprise and the API. Data is encrypted at rest and in transit. The privacy page does not mention a processing location. Another OpenAI page describes at-rest data residency for Enterprise, Edu and some API access, with a “Europe (EEA + Switzerland)” region; it does not mention Business. This concerns storage: we cannot infer that no data leaves Switzerland. Ask for it in the contract.
Personal accounts. The “Improve the model for everyone” setting decides whether your new conversations are used for training. The help page does not state its default value: open your settings and look. Temporary chats are not used for training and are deleted from OpenAI’s systems within 30 days.
And the nFADP? If Bexio holds data about clients, employees or third parties, Swiss law applies. Three reference points, taken from the text of the Act: data security must be adequate to the risk (art. 8), processing by a third party requires a contract or a legal basis, and no duty of secrecy may prohibit it (art. 9), and disclosure abroad requires an adequate level of protection or contractual safeguards (art. 16). For an accounting firm, professional confidentiality comes on top. A free or personal plan, with no contract, is rarely the right framework for real accounts.
Key takeaway: the question is not “is ChatGPT safe?”, but “which plan, which contract, which setting?”. Until you can answer all three, work on anonymised data or a test account.
The risks specific to ChatGPT
The token risks (leak, expiry, quota) are those of the article on Claude. Here are the ones ChatGPT adds or makes more visible.
- The token in a configuration. In a GPT, a plugin or a server, it is kept somewhere you do not always control. A single shared access is enough to expose it.
- Prompt injection. A document, an email or a Bexio field can contain hidden instructions. OpenAI warns that a server receiving untrusted content is a danger, even if its developer is trustworthy. A client name or an invoice note coming from outside is a possible way in.
- Unwanted writing. ChatGPT asks for manual confirmation before a write, but only tools carrying the “read-only” annotation are treated as such, and OpenAI warns that a server can mislabel its tools. Approving a tool “for the conversation” applies to the whole conversation: grant that only to a tool you trust.
- Writing that cannot be undone. An invoice that has been sent or an entry that has been posted cannot always be reversed with one click.
- A server that changes. According to OpenAI, a published plugin uses a frozen snapshot of the tools. Changes to the server only take effect after a new publication, and an incompatible change can make calls fail without warning anyone.
ChatGPT or Claude: what really differs
Without disparaging anyone: in substance, it is the same situation. In both cases, the Bexio token is the same, with the same rights, the same 60 days and the same API limits. In both cases, the AI provider states that it does not verify third-party servers. The differences are practical.
- The administration framework. At OpenAI, the ability to plug in a custom server depends heavily on the plan and the managed workspace, with detailed permissions. Read your workspace’s policy before you start.
- Stability. The move from GPTs to plugins is under way. An integration built today on an old route will have to be redone. With Claude too, access conditions evolve: date your checks.
- Token storage. With a tool installed on your computer, you can keep the token in an environment variable, as we detail in the article on Claude. With a ChatGPT plugin, it generally lives on a remote server. OpenAI also describes local MCP apps inside plugins, which run on your computer with the ChatGPT Desktop app (not on web or mobile). We could not confirm how a token can be kept there.
Neither option is “the right one”. Choose the one whose contract, data setting and token storage you understand.
What you can do with it, if you go ahead
Same rules as for Claude: start with reading. List invoices open for more than 60 days, spot duplicate contacts, summarise a quarter’s expenses, prepare a check before the VAT return. Writing (creating a contact, preparing a draft invoice) comes only after trials, with your confirmation each time. Posting entries without review, changing invoices that have been sent or touching salaries and closed periods: avoid.
All of this assumes clean accounts to begin with. Our guide Bexio for the self-employed sets out what the tool does well, and what it does not do for you.
Free, no-obligation first conversation. Request a quote
Six things to remember
- Check your plan first and your workspace policy: does the “custom MCP server” option exist for you, and who can switch it on?
- Do not build on a custom GPT: retirement announced for 11 December 2026, Actions not transferable.
- Treat the token like a password: 60 days, a reminder on day 55, never in a shared GPT or in a conversation, and a single server you trust.
- Control the data: training setting for a personal account, contract and DPA for professional use, processing location to ask for.
- Stay read-only until several answers have been checked by hand, then read the parameters of each write before confirming.
- Delete the token in Bexio’s developer portal, and disconnect the plugin, as soon as you no longer need it or if in doubt.
We can help you set it up
Connecting Bexio to ChatGPT is mostly about choices: which plan, which server, which storage, which uses, which limits. We work in Bexio every day (our Bexio service) and can help you set up the connection and define use cases that make sense for your business, deciding with you what may be written in your accounts. The first conversation is free and without obligation, and we can hold it in French or in English. Official filings with the Geneva authorities, however, are made in French.
This article gives general guidance, as at 27 October 2026. OpenAI’s offers and Bexio’s access conditions may change: check them before you start. For a specific situation, let’s talk during a first conversation.