A hands-on walkthrough for enabling MCP Server on Autonomous AI Database Serverless, registering your first tools, and connecting an AI agent client — the governed way.

In our last post, we introduced the Oracle Autonomous AI Database MCP Server — a built-in, fully managed feature that lets AI agents like Claude Desktop or OCI AI Agent interact directly with your database through the Model Context Protocol. Now let's actually turn it on and use it.

This walkthrough covers enabling MCP Server on Autonomous AI Database Serverless, registering your first tools, and connecting an MCP client. If you're running Dedicated Infrastructure (ADB-D), the flow is similar but goes through Database Tools instead of a free-form tag — we'll note the difference at the end.

You'll need:

• An Autonomous AI Database instance on 19c or 26ai
• An OCI IAM user with permissions to update the database
• An MCP-compatible client — Claude Desktop, OCI AI Agent, or VS Code with Cline all work
• Basic familiarity with the OCI Console and Select AI Agent (Oracle's framework for defining database-callable tools)

Here's the full path from client to database, and where the governance actually happens:

Unlike most OCI features, you don't flip this on through a settings panel — you enable it with a free-form tag on the database resource.

1. Sign in to the OCI Console as a user with permission to update your Autonomous Database.
2. Navigate to your Autonomous AI Database instance.
3. Under tags, add the following free-form tag:

Tag Name:  adb$feature
Tag Value: {"name":"mcp_server","enable":true}

Save the change. Adding this tag creates a dedicated MCP endpoint bound to your database's OCID. You don't need to provision anything separately — Oracle handles the endpoint creation automatically once the tag is applied.

Once enabled, your database exposes an endpoint in this format:

https://dataaccess.adb.<region-identifier>.oraclecloudapps.com/adb/mcp/v1/databases/<database-ocid>

Replace:

• <region-identifier> with your database's OCI region identifier (e.g., ap-mumbai-1)
• <database-ocid> with your Autonomous Database's OCID, found on the database's detail page in the OCI Console

Keep this URL handy — you'll need it when configuring your MCP client in Step 4.

The MCP Server doesn't expose your whole database by default — it exposes tools that you define and register through Select AI Agent. This is the layer that actually decides what an AI agent is allowed to do.

Common starting tools include:

• A schema discovery tool, so the agent can see which schemas/tables it has access to
• A SQL execution tool, scoped to a specific role, for running ad-hoc queries
• Custom business-logic tools specific to your application (e.g., “look up order status,” “generate test dataset for schema X”)

Now point an MCP-compatible client at your new endpoint.

Example: Claude Desktop

1. Open Claude Desktop's configuration.
2. Add a new MCP server entry pointing to your endpoint URL from Step 2, using the streamable-HTTP transport.
3. Save and restart the application — MCP client configuration changes typically require a restart to take effect.

Example: VS Code with Cline

1. Open Cline's MCP server settings.
2. Add the same endpoint URL under a new server configuration.
3. Reload the extension.

Once connected, your client should list the tools you registered in Step 3. You can now ask your AI agent things like “show me the schema for the orders table” or “run a query to check today’s enrollment count” — and it will execute through the governed, audited path you configured, not an unrestricted connection.

After your first few tool calls, check that everything is behaving as expected:

• Confirm activity is showing up in OCI Monitoring and your database's audit logs — MCP Server activity is captured through the same auditing and monitoring tooling as any other database access.
• Watch your ECPU usage, since queries run by an AI agent still consume compute like any other workload. Autonomous Database's auto-scaling helps absorb spikes, but it's worth keeping an eye on if agent usage grows.
• Test what happens when the agent tries something outside its granted role — it should be denied, the same way a human user without permissions would be.

The Dedicated Infrastructure flow is a bit different — here's how the two compare side by side:

Instead of a free-form tag, you create the MCP Server explicitly under Developer Services → Database Tools → MCP Servers in the OCI Console, tied to a Database Tools Connection (backed by an OCI Vault for credentials) and an OCI Identity Domain in the same region as your database. Authentication uses OAuth 2.0 with PKCE rather than the simpler tag-based flow.

This is just the entry point. From here, the natural next steps are building custom Select AI Agent tools tailored to your application, tightening access with role-based scoping, and integrating MCP-driven workflows into real development pipelines — schema-aware code generation, automated migration drafting, and AI-assisted troubleshooting among them.