How do I isolate Pydantic AI Montage authorization by user session?
A Pydantic AI toolset can wrap an authorized FastMCP client, but it must not become a vehicle for sharing one user’s workspace access with others. Review the client’s account association before exposing it through Agent toolsets or starting a model-driven research task. The configured FastMCP Client must remain associated with its requesting user when passed into MCPToolset. A technically valid toolset can still cross an account boundary if the application reuses private authorization indiscriminately across sessions.
Help me isolate Pydantic AI Montage authorization by user session. Determine which application user and request should own the Montage client. Report the checked session’s actual response and any unresolved association risk. No writes.Identify the session owner
Determine which application user and request should own the Montage client. Keep this association in application policy rather than allowing model-generated arguments to select another account. The requesting-user association should be documented without including real account identifiers in shared prompts or using model-supplied values as the authority for client selection.
Trace client injection
Review the documented path that passes a configured FastMCP Client into MCPToolset. A URL-only construction cannot identify the user or establish approved authorization by itself. The supported client-injection path is the relevant integration contract; a bare remote URL cannot stand in for the configured FastMCP authorization owner.
Inspect private auth storage
Check supported callback and token handling against the application’s user boundary. If clients or auth state are shared ambiguously, fail closed before attaching private workspace tools to an Agent. If storage isolation is ambiguous, that ambiguity is a blocker even when the server accepts the shared grant and technically returns valid data.
Read within the intended session
Execute signals_status directly with that authorized client and no Agent run. Compare the response with the expected workspace without dumping private storage or another user’s account details. The direct response should be compared with the intended user context before enabling any broader toolset, without exploring an unexpected workspace’s retained records.
Limit the isolation claim
Report the checked session’s actual response and any unresolved association risk. One correct workspace result does not prove every multi-user path is isolated or authorize reusing the client for unrelated people. Isolation evidence must remain limited to the checked path, with other users and concurrent requests identified as separate unverified cases rather than implied successes.
Keep the boundaries clear.
- If storage isolation is ambiguous, that ambiguity is a blocker even when the server accepts the shared grant and technically returns valid data.
- Isolation evidence must remain limited to the checked path, with other users and concurrent requests identified as separate unverified cases rather than implied successes.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.