Your accountant needs to explain why a balance differs, not write a prompt without the underlying records. Copying transaction files into a chat loses context: which entity owns the account, which period you are reviewing and which rules apply.
MCP connects an AI assistant to the tools a software product exposes. Conta Finance uses it to let an authorized assistant consult an organization’s records and prepare changes. The person reviews and confirms those changes in Conta Finance. The model does not get the final decision.
What is MCP, in practical terms?
Model Context Protocol is an open standard for connecting AI applications to external data and tools. A client requests a tool, and a server returns the result. MCP is the connection protocol, not an accounting model or a guarantee that an answer is correct.
For a finance team, the useful question is not “Does this product have AI?” It is “Can the assistant retrieve the right records, within the right entity, without making unauthorized changes?” A fluent answer is not enough if you cannot check its source.
ContaGPT and MCP are two different ways to work
ContaGPT is the assistant inside Conta Finance. With MCP, you work from an external assistant’s own model. It calls the tools Conta exposes; it does not ask ContaGPT to answer on its behalf. Your organization remains the source of the records.
This suits teams that already use an AI workspace. Instead of moving every investigation into a new chat, they can bring relevant accounting context into their existing workflow. Each connection is for one organization, so selecting a company remains an explicit choice.
Three questions to test in a finance workflow
Start with a bounded task and a period. Ask for the underlying records, not only a summary. For example, investigate one account before asking the assistant to explain an entire group’s activity.
- “Which transactions in this account still need review? Show the references so I can open them.”
- “Explain the balance difference for this token. Distinguish missing activity from a pricing issue.”
- “Propose a category rule for recurring validator rewards. Show what you would change.”
A useful answer must be verifiable
Review the account, dates, amounts and linked records. If the available data cannot explain the difference, keep it as an open question. Do not turn an AI inference into a booked fact.
How Conta keeps access and changes under control
A person with full access enables AI clients for the organization. The user then authorizes a connection through OAuth. The connection’s permissions can narrow the user’s access, but cannot expand it. A read-only user can consult records, not prepare changes.
An assistant can prepare a proposal when the user’s access allows it. It receives a link to the app. The person opens that link, reviews the proposed change and confirms or discards it. The MCP client cannot confirm the proposal, even if its own interface offers automatic approval.
- AI assistantPrepares a proposal within your permissions
- Conta FinanceYou open the link to the pending change
- Your reviewYou check what would change
- Your decisionConfirm or discard
The assistant cannot confirm a change. You decide inside Conta Finance.
What to check before connecting an AI provider
An authenticated connection is not a promise that information never leaves Conta Finance. The external assistant receives the results it requests. Your team must review its provider’s retention, training, residency and sharing policies, as it would for another export of financial information.
Define who can connect an assistant, which organization they can access and which questions are appropriate. Use a limited test scope first. A model can misunderstand a transaction or follow misleading third-party text; authorization and human review remain necessary.
How to assess MCP in a Conta Finance demo
Use an assistant that supports authenticated custom MCP connectors. Compatibility depends on the client, its application and its plan. Claude has been tested as a custom connector in our development environment. Do not assume that every assistant or subscription supports the same connection.
Bring a specific question about a balance, a transaction or a rule. Ask us to show the returned evidence, a read-only account and a proposal that requires confirmation. These checks tell you more than a general demonstration of conversational AI.
Frequently asked questions
Can an AI assistant change the books without approval?
Not through Conta Finance MCP. It can prepare authorized proposals, but a person confirms them inside the application. Read-only users cannot prepare proposals.
Does MCP replace the accountant?
No. It supports retrieval and investigation. The accounting team still checks the evidence and decides the treatment. An AI answer is not an accounting policy.
Does the external assistant use ContaGPT?
No. Its own model uses the tools Conta exposes. ContaGPT is the separate assistant inside the application.
