Skip to main content
Connect agents to external tools and services using the Model Context Protocol. MCP (Model Context Protocol) is an open standard that lets AI agents interact with external tools and services through a single, consistent interface. Instead of building a custom integration for each tool, agents connect to MCP servers and access all the tools they expose through one standardized protocol. Learn More. Without MCP: Connecting to multiple external services requires a separate integration with custom logic for each one. With MCP: The agent communicates with all services through a single interface, dramatically reducing development complexity.

When to use

MCP is particularly suited for tools hosted on external servers, shared toolsets across teams, and cases where you want clear separation between tool logic and agent logic.

How MCP works

MCP uses a client-server architecture with three components:

Interaction workflow

  1. Tool Discovery - Agent Platform MCP client connects to the MCP server and retrieves the list of available tools.
  2. Intent Detection - The LLM receives the user query and the tool list, then identifies the appropriate tool.
  3. Tool Invocation - The client sends a structured request with the tool name and required parameters.
  4. Execution - The server runs the tool and returns results to the client.
  5. Response Generation - The agent uses the tool output to formulate a natural language response.

Example

Configure MCP server

Navigate to Tools > MCP Servers, provide the following and Register Server. Do the following after successfully registering an MCP server:
  • Configure connection settings.
  • Test server connectivity.
  • View discovered tools.
  • Import tools into the platform.

Configure connection settings

Dynamic URLs and parameters An MCP server URL can use {{env.<name>}}, {{config.<name>}}, and {{input.<name>}} references. Custom headers can use {{input.<name>}} references. In the server configuration, choose Parse to find input references in the URL and headers and add them to the server’s parameter list. Add environment and config variables in their respective variable settings; they are resolved at run time and are not added as input parameters. You can also add input parameters manually.
  • Server parameters can have default values. If a tool call omits a value, the configured default is used for URL and header placeholders.
  • Every parameter used in the URL must have a default. Studio marks its default field as required and prevents saving the server without one.
  • Studio uses the default URL for tool discovery, import, and connection testing. It shows this URL on the Configuration and Tools tabs. For example, a URL of https://{{input.region}}.example.com/mcp with region defaulted to us uses https://us.example.com/mcp for those actions.
  • At run time, the platform resolves {{env.<name>}} and {{config.<name>}} references using the session’s environment before substituting {{input.<name>}} values. For example, https://{{env.MCP_HOST}}/{{config.API_VERSION}}/{{input.region}}/mcp resolves to https://mcp.example.com/v1/eu/mcp when the session resolves MCP_HOST to mcp.example.com, API_VERSION to v1, and the call supplies eu for region. This works for {{env.<name>}} in a server URL even when the server does not link a variable namespace.
  • Input values are inserted after environment and config references are resolved. A caller-supplied value does not get interpreted as another {{env.<name>}} or {{config.<name>}} reference. The final URL is checked before the platform connects; a URL that fails the safety check is rejected.
  • If a call supplies no value and no default is available, the call fails.
When Connection reuse is enabled, calls to the same resolved URL in a conversation can reuse that URL’s MCP session. Calls to different resolved URLs use separate sessions. This also lets an agent and workflows it starts reuse the session for the same URL.

Test connection

Click Test to verify server connectivity and validate the current configuration. This confirms server reachability, authentication validity, transport configuration, and request handling.

Discover and import tools

After registering a server, the platform automatically discovers the tools it exposes. You can then:
  • Click Import next to a specific tool to add it individually
  • Click Import All to add all discovered tools at once
Once imported, tools can be attached to one or more agents.

View and test tools

Click on a tool to view its description, metadata, input and output parameters and to test the tool. Go to Testing and select Run Test, choose the environment (Dev, Staging, or Production) under Run context, enter sample values for the input fields, and verify the response. Any environment variables in the tool’s headers use the values of the environment you choose.
Custom header precedence: Headers configured on the MCP server apply to all its requests. A header configured on an individual tool binding overrides the server header with the same name. Header names are matched case-insensitively, so X-API-Key and x-api-key are the same header. Headers present in only one configuration are retained. For example, if the server sets X-API-Key: shared-value and the tool sets x-api-key: tool-value, the request uses tool-value. The server and tool settings remain separate in configuration; they are combined when the request is prepared.

Refresh MCP server

Use the Refresh icon to rediscover tools from the server. Do this when new tools are added, existing tools are updated, or tool metadata changes. Tool definitions are fetched at configuration time and aren’t automatically synced.

Attach MCP tools to an agent

Go to the selected agent, then navigate to Tools > Attach Tool. Select tools from the list. MCP tool names are prefixed with the MCP server name when attached to an agent. You can also attach tools via the ABL file using the /mcp-tool insert command:
where,
  • server - MCP server name
  • tool - Tool name on the MCP server