MCP prompts are reusable, server-defined message templates that resolve into a sequence of chat messages. You can use prompts to standardize how the model receives instructions and reuse that pattern across multiple agents (your MCP server supports Handlebars placeholders like {{argumentName}}).
Adding an MCP prompt to an agent lets you:
-
Reuse proven instruction patterns instead of rewriting them per agent.
-
Improve consistency by centralizing prompt templates on the MCP server.
-
Reduce prompt drift by keeping the structure stable and only varying inputs through arguments.
-
Separate concerns between prompts (instructions and formatting) and tools (actions and data retrieval).
When you add an MCP prompt to an agent:
-
The agent discovers and references the prompt through the MCP connection.
-
At runtime, the agent renders the prompt by filling in the prompt arguments.
-
The rendered output becomes a message sequence that instructs the model.
The agent runtime supplies argument values, not the user. Depending on your configuration, argument values come from workflow context and variables, outputs from prior steps, or other runtime inputs. Prompts provide structured instructions; they do not execute tools or change external systems.
-
Navigate to AI studio → AI agents → Create AI agent.
-
You can select one of the following alternative options: click + Create Agent or click an existing agent or click (...) in the Actions column and select Edit AI agent.
-
Go to the Tools & capabilities section and click + Add.
-
In the Add capabilities page, on the top of the page select MCP connections.
-
In MCP connections, you can either:
-
On the top right, you can either click + Create an MCP connection to create a new connection, or
-
Select an existing MCP server
For setup details, see MCP server.
-
-
Expand the connection to view available tools and prompts.
-
Select the prompts to add to the agent and click Save.
The selected prompts are now available to the agent.
Once you add prompts to an agent:
-
The agent uses the standardized instruction template as needed.
-
The agent supplies values for the prompt's arguments from its available runtime context.
-
The MCP server returns the prompt as a rendered message sequence (after argument substitution), which the agent uses to guide the model.
Prompt does not appear in the selection list
-
Confirm the prompt exists in the MCP server's → Capabilities → Add → Prompts tab.
-
Confirm the agent uses the correct MCP connection.
-
Refresh and expand the MCP connection to reload tools and prompts.
Rendered output shows placeholders
-
Verify the placeholder name matches the argument name exactly. For example, {{customer_name}} in the template requires an argument named 'customer_name' and not 'customerName'.
-
Ensure the agent runtime supplies a value for that argument.
-
Use a stable name as the prompt identifier.
-
Write the title and description so builders can choose the correct prompt quickly.
-
Keep templates explicit (expected output shape, constraints, tone) and keep variability in arguments.
-
Do not place secrets in prompt templates or argument values without explicit secret handling.