Questions IT teams ask about AI agent access

MarcoPolo, built by Immersa, Inc., is a data gateway for AI agents: one governed MCP connection point between AI assistants or custom agents and your databases, warehouses, cloud storage and SaaS applications. Credentials stay in MarcoPolo, data work runs in an isolated workspace outside the model, and every operation through the gateway is recorded in an activity log. Read in the docs

Add MarcoPolo once in each assistant: in ChatGPT from the Apps directory, in Claude as a custom connector. Set up each data connection once in MarcoPolo, where its credentials stay. Each person signs in as themselves and the assistant can use only the connections shared with them. Queries run against the source systems without copying the data, and every operation lands in one activity log for both assistants. Read in the docs

Yes, and it goes further than routing. Assistants connect to one MarcoPolo MCP server instead of one server per system. MarcoPolo decides which connections each identity can see, injects credentials in a privileged layer the model cannot reach, and logs each operation. Each identity also gets a persistent workspace with DuckDB, Python and a shell, so query results are cached and analyzed there and only summaries return to the model. The same gateway is available through a Python SDK and a REST API. Read in the docs

Set up each connection once in the MarcoPolo web app. Its credentials are encrypted and kept in a secure store, never in a workspace. Employees connect their assistant to MarcoPolo over MCP and sign in as themselves. The assistant can use the connections that person owns or that were shared with them or their organization, and it never receives the keys. Read in the docs

Any MCP-compatible client, including Claude, ChatGPT, Cursor, Codex, VS Code, Replit and custom agents. In ChatGPT, MarcoPolo is listed in the Apps directory (Settings, then Apps, then search for MarcoPolo). In Claude, you add it as a custom connector, and plugins for Claude and Codex add guided data workflows. Read in the docs

No. The AI works in user mode inside the workspace. A separate privileged execution layer retrieves the encrypted credentials, runs the query and returns only results. Decrypted credentials are never written to the workspace, passed to the model, included in conversation history or logs, or shared between users. Read in the docs

Every call carries a credential that selects one identity: a person, a service account or an application user. That identity determines the workspace, the visible connections, the permissions and the audit attribution. A connection is visible only to its owner and to the people, service accounts or organization it is shared with, and each user works in a separate, isolated container. Read in the docs

The Activity page records data operations performed through the gateway: when each one ran, the identity and connection involved, the action and its outcome, and, depending on the operation, the MCP tool, the query text, the row count and the duration. Users see their own activity; administrators review activity across the organization and export it. Events are stored in OpenTelemetry logging and tracing services. Read in the docs

MarcoPolo does not replicate or warehouse your data. Queries run against the systems that own it. The results a user pulls are cached in that user's workspace for follow-up analysis, and customer data is never used to train AI models. Read in the docs

Yes. Use the hosted service at mcp.marcopolo.dev, or deploy MarcoPolo in your own AWS, Google Cloud or Azure environment, or on premises on VMware or OpenShift (Bring Your Own Cloud). Read in the docs

Through the Python SDK, the REST API or MCP. An application can run as a service account with explicitly granted connections, or exchange a namespace key for short-lived tokens so each application user gets their own workspace, permissions and audit identity. Read in the docs

Individual developers can start free with the getting started guide and a demo connection, which skips the credentials step. For a team or enterprise rollout, book a walkthrough with the MarcoPolo team. Read in the docs