How MCP works
An MCP client connects to a server and discovers its available tools. Each tool has a name and an input schema describing the arguments it accepts. The client can then call a tool and receive the result. MCP also supports resources for reading data and prompts provided by the server.
Advertising workflows an agent can help with
An operator might ask an agent to review underdelivering campaigns, query revenue by placement, or prepare a line item. Tools give the agent a documented way to read and change platform configuration. Access permissions, approval rules, and the available tools determine how far it can take each task.
Control access before enabling changes
- Give each agent an identity and access to the tenants it is authorized to serve.
- Limit the operations, spend, and level of autonomy it can use.
- Require human approval for selected changes. Permissions and spending caps still apply after approval.
- Record changes and refused actions so operators can review what happened.
Riptide provides these controls through its agent gateway. Registered agents prove their
identity with signed requests. MCP clients connect over HTTP at POST /mcp
and can use tools, resources, prompts, and approval tasks. REST tool discovery and
invocation are available under /v1/tools.
Start with the agent controls overview, browse example MCP tools, then use the connection guide and downloadable clients to test access in your tenant.