External tools (MCP)
Connect standard remote MCP servers to a custom model and manage authentication, tool discovery, and timeouts.
Model Context Protocol (MCP) lets a custom model use standard tools exposed by a remote service. UnieAI Studio connects to a remote MCP server, performs initialize and tools/list to discover its tools, and then makes eligible tools available to the agent.
External tools (MCP) connects only to standard MCP servers. Knowledge bases, web search, SQL, API tools, and skills have separate configuration pages. Studio does not automatically convert an arbitrary HTTP API into an MCP server.
Enable remote MCP
Open the custom model's Customize page and go to Behavior → External tools (MCP).
Turn on Enable MCP tools, then select Add MCP server.
Enter the server name, URL, authentication method, and any required headers.
Select Verify & list tools. Check the discovered tools before saving the server.
Make sure Active in runtime is on and save the custom-model configuration.
An inactive server remains saved but is excluded from the agent's toolset.
Server settings
| Setting | Meaning |
|---|---|
| Name | A label used to identify the server in Studio |
| Server URL | The complete HTTPS URL of the remote MCP endpoint |
| Auth type | None, bearer token, or OAuth 2.0 |
| Additional headers | A JSON object for tenant identifiers or provider-required headers |
| Tool timeout | Use the CM default timeout is checked by default. Uncheck it to give this server its own timeout, from 1 to 600 seconds |
| Active in runtime | Whether this server contributes tools to the agent |
Bearer token
Choose Bearer token and enter the token. The runtime includes the corresponding authorization information when calling the server. Grant only the permissions needed for the task and rotate the token according to the provider's policy.
OAuth 2.0
Choose OAuth 2.0, optionally enter a client ID, client secret, and scopes, and select Connect via OAuth. Studio attempts to discover the authorization endpoints. If the server supports dynamic client registration, some client details can be created automatically. After authorization, the access token is used for MCP calls and a refresh token is used when available.
Reconnect if the token expires and cannot be refreshed. Save the server again after removing scopes or disconnecting OAuth.
Additional headers
Headers must be a valid JSON object:
{
"X-Tenant-ID": "tenant-123",
"X-Environment": "production"
}Do not place the same long-lived secret in both the bearer-token field and a custom header.
Verification and tool discovery
Verify & list tools connects to the MCP server, initializes a session, and requests tools/list. Before saving, confirm that:
- The URL points to the intended environment
- The account and OAuth scopes are appropriate
- Tool names, descriptions, and input schemas are correct
- Administrative or destructive tools are not exposed unnecessarily
A successful verification confirms connectivity at that moment; it does not guarantee the provider's future availability or output.
How the agent calls MCP tools
Agent Core 1.0 converts discovered tools from active servers into native function-calling tools. Runtime names use the mcp__ prefix. When the model chooses one, the runtime validates its arguments, applies authentication and timeout settings, calls the MCP server, and returns the result to the model.
Timeout precedence is: server-specific timeout, custom-model MCP default, then the runtime default. Agent Core also applies overall step, repetition, and per-tool execution limits.
The maximum-call setting is an MCP selection cap for flows that consume it. Do not treat it as the only security or cost boundary. The native Agent Core loop is also governed by AI Engine step limits and runtime safeguards.
Import and export
You can import MCP server JSON or export a saved configuration for another environment. After importing, verify the URL, tool list, and authorization scope again.
Exported JSON may contain bearer tokens, OAuth access or refresh tokens, client secrets, and custom headers. Treat it as confidential. Do not commit it to Git, paste it into a ticket, or share it through a public link. Delete it securely after use and rotate credentials when necessary.
Security recommendations
- Prefer trusted MCP servers over HTTPS.
- Use a dedicated service account and minimum scopes instead of a personal administrator account.
- Treat tools that write, delete, pay, or send data externally as high risk.
- Review tool descriptions and returned content; external content can be incorrect or contain prompt-injection text.
- The MCP server receives arguments supplied for the tool call. Do not put unrelated secrets in the prompt.
- Repeat discovery and Playground testing after changing a server, scope, or tool version.