Skip to main content
An MCP server describes tools exposed by an external service. Review the server and each intended action separately: knowing an endpoint does not tell you which data its tools can access or change.
Server configuration can be saved, but Test Connection is disabled and server tool discovery returns an unavailable result. This guide does not establish live discovery, runtime approval enforcement, or MCP tool execution.

Prepare the server and tools

1

Review and save the server

Follow MCP servers to review its operator, endpoint, transport, and credential reference. Record the resulting server ID.
2

Document the intended actions

Obtain the tool names, descriptions, and input schemas from the server operator. Identify read and write actions separately. Keep discovery verification pending; do not present a manually copied schema as a successful connection test.
3

Review approval intent

The tool contract includes approval_policy values require_approval_all, require_approval_per_tool, and auto_approve_all. Decide which actions should require confirmation, but verify enforcement separately when execution is released.
4

Inspect agent references

Custom server references belong in conversation_config.agent.prompt.mcp_server_ids. Preserve the intended list in a complete agent draft and read it back. Selecting a server in the dashboard’s Tools > MCP list currently opens its details; it does not save an attachment.

Tool-level contract

The MCP tool shape includes type: "mcp", integration_type, mcp_server_id, mcp_server_name, mcp_tool_name, and mcp_tool_description. It can also describe parameters, mcp_input_schema, and input_overrides. These fields describe a tool configuration. They do not make tool-management endpoints available to personal API keys, and they do not create a functioning server connection. Use the tool overview for the current management restriction.

Frequently asked questions

Not by itself. A stored policy is authoring intent. Runtime confirmation and enforcement need separate verification when the execution path is available.
The agent schema separates mcp_server_ids from native_mcp_server_ids. Use the identifier required by the relevant integration workflow, rather than moving IDs between these lists.