Skip to content
Interface language: English
Audit / Signal

An MCP server changed its tools: is it still the integration you approved?

An MCP server can keep its address while its tools change. Compare the catalogue, descriptions and parameters and decide which changes need renewed approval.

Article language: English

A woman in a red apron beside a round glass lens in a gold-coloured housing.
In this article

On Monday an integration could find a message and prepare a draft reply. On Thursday, at the same server address, a tool appeared that sends a reply to everyone in the thread. The provider name had not changed, the icon was still green and the register said “approved Monday”. Only the range of possible actions had grown a little, like a plant someone had been watering with permissions.

This is a fictional example, but the acceptance question is specific: what was approved? A connection to the server, mailbox reading, a list of particular tools, or every action the provider might later add? If the approval record cannot distinguish these, checking that the address still responds will not help by itself.

MCP is a protocol that lets an AI application use tools exposed by a server, among other capabilities. Reviewing the catalogue in an operating process should reveal what changed since the last approval and how that difference affects the task. The process owner then decides whether to continue using the integration. The trial below covers tool behaviour and its operational acceptance.

The same address does not describe the whole action

In our mail process, the approved activity was finding information and preparing material for human review. Sending was outside scope. Adding a separate sending tool does not automatically mean it will be used, but it changes the set of actions the application can be offered. Check whether the application still stops at a draft after the change.

The MCP tools specification retrieves a catalogue through tools/list, and a tool definition includes a name, description and input schema among other elements. listChanged concerns notification that the list changed. The notification itself is not an assessment of the new scope of action.

To compare two states, keep more than a screenshot of the connection icon. Preserve the server identifier, actually exposed tools, their descriptions and parameters, and the account used to retrieve the catalogue. A test account with limited access may see a different scope from the team's connection. Do not present that read as evidence for every account.

A tool name helps find an item but does not replace assessment of its action. A new item might clearly indicate sending. A harder case retains the old name while changing a default: preparing a draft becomes sending a reply when an additional parameter is omitted. Comparing names alone will not show that difference.

When combining integrations, also record which connection supplied each item. Similarly named tools may refer to different mailboxes or environments. “Search checked” settles nothing if it is unclear whether it meant a test catalogue or company mail. The identifier should lead back to the assessed configuration without guessing from a logo colour.

The article on AI integration permissions (in Polish) describes the basis for allowing data and actions. Here the additional issue is change over time: does the later catalogue still fit agreed boundaries? Reconnecting, refreshing the list or signing in again should not itself substitute for answering that question.

Record three separate approval findings:

  • Task: prepare a draft from a specified message, with human review before sending.
  • Access: the particular mailbox and materials needed for the answer, with an identifiable account used by the process.
  • Action: tools permitted for that task and actions outside scope, even if the server can perform them.

This permits assessment without debating whether someone once “agreed to MCP”. The protocol is a connection method. The decision concerns particular use of data and execution of actions.

A change can hide in a description or parameter

Compare the approved state with the present one. Separately identify added, removed and changed tools. For a changed item also show the description, input and meaning of the result. A shorter description need not mean a simpler action; a new optional parameter can change the default behaviour of every existing call.

Consider three changes in our mail example. A sending tool is added. An existing tool gains a recipient choice defaulting to the entire thread. Or its description begins encouraging the model to send finished material after finding an answer. Each needs a different check: availability of the new function, parameter meaning and how the application selects an action.

For a parameter test an omitted value, an empty value and an explicitly supplied value. If the approved trial covered only a particular recipient list, it does not establish behaviour without the list. Record what was actually sent to the tool and what happened in the agreed environment. The model's “I prepared a draft” does not prove the message remained a draft.

The meaning of the result can change too. A field that previously identified a draft might after an update identify a sent message. Syntax validation still passes if both values are text. The reviewer needs the field's meaning in that version and must compare it with the actual case state. A neat data object can be a very polite wrapper for the wrong outcome.

Aeneia conceptual photography

The official MCP blog on tool annotations stresses that they are hints, not guarantees of behaviour or mechanisms enforcing boundaries. readOnlyHint alone does not block writes. Separate descriptive information from controls that actually restrict execution.

Treat “safe search” similarly. Ask which resource is searched, what leaves the environment and whether the call can perform an additional action. A description should help explain the function. It should not itself grant permissions or establish that a new scope has been accepted.

Catalogue checking complements control of foreign instructions found in documents. Prompt injection in company material (in Polish) concerns source content attempting to direct a model's actions. With tools, assess the function description supplied to the application and actual execution capabilities as well. These paths can meet, but require separate evidence.

Record the catalogue-check time and action-execution time separately. The team may have read a catalogue in the morning, while a later call uses a different state. Show which moment the comparison covers and whether the trial actually used that set. If that cannot be reconstructed, state the evidence limit. Do not present a morning read as confirmation of evening behaviour.

Agree what happens when a difference is detected

Not every change requires stopping the whole process. A corrected typo may not change behaviour, while a new write location or sending to additional recipients may exceed approved scope. Record the decision criterion and decision-maker. “Minor change” should identify the effect checked, not the number of characters changed.

For a new function outside scope, establish a waiting state: the tool remains unavailable to the process pending a decision, or the whole integration enters an agreed restricted mode. The choice depends on the solution's capabilities. In the trial, check the actual result of that setting, including after reconnection and catalogue refresh.

Removing an item raises a different question. Does the task end with an explicit absence, or does the model try to substitute a broader tool? After losing a draft function it should not select sending simply because both descriptions mention replying to a message. A substitute must fit the task and agreed permissions.

If a change fails acceptance, identify a real way to limit its effect. Restoring an earlier configuration may stop future sends but cannot recall a message already sent. Separate returning to the previous way of working from addressing actions performed under the changed scope. In the mail trial check whether a queue of pending sends remains after blocking. “Integration disabled” alone does not explain that queue's state.

An AI agent's sandbox and file and network access (in Polish) discusses execution boundaries. A sandbox is an isolated environment with specified access to resources. Here, check whether the restriction also covers a route through an external server. A local write prohibition does not yet explain what a tool in another system can do.

The same applies to excluding files from a coding assistant's work (in Polish). If a new function can read a repository through a separate service, an old local directory filter may not describe that route. Compare both connections' scopes. Do not write “file unavailable” solely because of a setting in one layer.

Leave evidence for a particular state

Plan the mail trial on a controlled account and messages prepared for acceptance, without unintended recipients. Preserve catalogue states before and after the change. Then use the same instruction that previously ended with a draft. Inspect the selected tool, parameters and mailbox result. This is a proposed control, not a test performed in this article.

Add a separate case omitting the new optional parameter. This shows whether the update preserved the old meaning of the call. Then check behaviour after refreshing the connection. If the application shows an old description but calls a changed action, interface evidence alone is incomplete. Distinguish the state visible to the user from what the integration executed.

If different accounts use the integration, check each role covered by acceptance separately. A preparer may see only a draft while an operator with broader access also sees sending. Do not count the first result as confirmation of the second. Record role, configuration and expected action. After the next update, the team can identify what to retest rather than reconstructing the history from memory.

For mail, also assess the result by the thread and actual recipients of an AI reply (in Polish). Correct text sent to the wrong people does not fulfil the task. A new “reply all” tool shows particularly well why the action name, parameters and effect belong in one acceptance record.

Place an identifiable configuration, the difference assessed, trial scope, decision and owner in the register. Do not keep secrets or an entire mailbox merely to prove one tool changed. A link to instruction versions and prompt-library owners (in Polish) helps reconstruct the instruction used in the trial, but does not replace the integration's own version.

For a conversation about Aeneia Audit, bring the earlier approval record, the present catalogue and one process where the change could matter. This permits a decision on what still works within agreed boundaries and what needs new approval. In our example the icon stayed green while the scope of action changed. The acceptance record should make the difference recognisable without sending real correspondence.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear