Connector
An adapter such as PostgreSQL, Redis or Kubernetes. It defines configuration fields, available tools and how calls are classified.
Getting started with GateRook
From your AI
to your infrastructure.
Connect an existing service, define what your AI may do, and make your first MCP call. One local gateway holds the connections and enforces your rules.
- 01
Your MCP client asks for context
An AI app connects to a GateRook endpoint and discovers the tools available there. A prompt such as “Find the stuck order” can become a database query and a log search.
- 02
GateRook checks the target and policy
The endpoint selects which connections and tools are exposed. Tool modes and connection policies decide whether the request can run, needs approval, or is denied. An approval does not override a denied policy.
- 03
The connector calls your service
GateRook resolves connection credentials and invokes the database, API, cloud service or host. For changes requiring approval, you review the exact action and target before execution.
- 04
Your AI receives the allowed result
The result returns to your client with masking and configured limits. Activity records the call and decision. Your AI uses the returned context to continue the conversation.
GateRook is the MCP gateway. Connectors are the adapters to your services; your AI app is the MCP client. You do not need a separate MCP server for each built-in integration.
Four things to understand
A small vocabulary. A clear boundary.
Connection
A configured instance of a connector: for example, a development database named dev-db. It holds the target, credentials, environment and policy.
Endpoint
An MCP entry point that exposes selected connections and tools. Use different endpoints for development and production, with access appropriate to each task.
Tool mode
Off disables a tool. Read only allows operations classified as reads. Approval requires your decision before execution. Not every tool supports every mode.
Your first endpoint
Connect once.
Keep the scope clear.
Before you start, have GateRook running, an MCP-capable AI client, and access to a service your computer can reach. Connect your VPN if your infrastructure requires it.
Choose a development resource for the first test. Use the service’s own permissions to grant only what the task needs.
Choose a connector and its setup guide- 01
Protect your master key
Follow the initial key-protection and backup flow in GateRook. Local secrets are stored encrypted. The master key belongs in your chosen secure storage, not in your AI conversation.
- 02
Add and test a connection
Open Connections, choose Add connection, and select the connector. Set a name and environment, enter its settings, and run Test connection. Complete the wizard to save it and choose its endpoint. Credentials entered here are resolved by GateRook.
- 03
Create or review your endpoint
In Endpoints, create an endpoint such as dev and select the connections it may use. In Tools & policies, enable the read tools you need. Leave unrelated tools Off. Enable an Approval mode only when you intend to review those actions.
- 04
Connect your AI app
In Clients, select your AI app and the endpoint. Review the connection method and any proposed configuration changes, then choose Connect. Follow the restart or authorization instructions shown for that client.
How client connection methods differ - 05
Make a small read in Playground
Select your endpoint, connection and tool in Playground. Try a schema listing or a small bounded query using the integration guide. Check the result, then ask your AI to perform the same task.
- 06
Inspect the call in Activity
Open Activity to see the requested tool, connection, decision and result. If you enable changes later, configure approval notifications and review requests in Approvals while the AI waits.
Bring your AI
One endpoint. Your choice of client.
GateRook includes guided flows for these AI apps. Use the settings generated by Clients so the transport, endpoint and authentication match your installation.
Claude Desktop
Use the desktop extension flow or the bridge configuration offered by GateRook. Review the installation request in the client. Keep GateRook running so the local bridge can reach the gateway.
Codex
Use the managed helper configuration in Clients. Review the proposed config change before connecting. The helper obtains its access credential when needed instead of placing a service password in the client config.
Cursor
Select the endpoint and a supported connection method in Clients. GateRook offers the relevant config review and authentication flow; restart or reload the client when instructed.
Antigravity
Choose the endpoint in Clients, review the generated connection details, and authorize access using the offered method. The endpoint controls which infrastructure tools become available.
Other MCP clients and remote clients
Copy the actual MCP URL from the endpoint. Its local shape is http://127.0.0.1:<core-port>/mcp/<endpoint-slug>; the port is installation-specific. Configure the transport and authentication supported by your client. Never substitute your database password for an endpoint credential.
A cloud-hosted client cannot reach your computer’s loopback address directly. Use a deliberately configured reachable endpoint, such as GateRook’s tunnel flow, and the corresponding authorization. A tunnel is optional for local clients.
Before a change runs
Review the action.
Make the decision.
When a call requires approval, GateRook holds execution. Review the target environment, connection and exact operation. An AI-written explanation is context, not a substitute for the command.
Approve only the action you intend, or deny it. Disabled tools and policy-blocked requests cannot run. Activity keeps the decision with the call.
Try the approval demoAn example to review dev
sql_execute · dev-db · illustrative request
The AI asks to cancel one order. The SQL is the action you review.
SQL
UPDATE orders
SET status = 'cancelled'
WHERE id = 48213;A real request shows the actual connection and operation. This documentation does not execute SQL.
When something does not connect
Check one boundary at a time.
Service access, endpoint policy and AI-client access are separate steps. The failure location tells you where to look.
Test connection fails
Check host, port, network or VPN, TLS settings and the service account’s permissions. Confirm the credentials are valid. Use the specific integration guide for required fields and supported authentication modes.
The connection works, but my AI sees no tools
Check that the connection belongs to the endpoint your AI uses and that the desired tools are enabled. Review endpoint authorization and reconnect or reload the client if its tool list is cached.
A query is denied
Inspect the call in Activity. A tool mode, connection policy, target restriction or service permission may block it. Check the exact tool and operation; a Read only mode does not authorize a statement classified as a mutation.
The AI is waiting for a result
Look in Approvals for a pending request. Check that GateRook is running, the service is reachable, and the request has not timed out. Review the exact action before deciding.
Does a local gateway mean all AI processing is local?
No. GateRook handles its configuration, policies and infrastructure calls locally. Allowed results go back to your AI client. A cloud model may process those results on its provider’s infrastructure. Choose your client and exposed data accordingly.
Your next step
Pick your first connection.
Start with a database, an API, your logs or a cloud resource.