How to Add a Custom Internal MCP Server to a Seahorse Cloud Agent

Seahorse Cloud managed agents support MCP-standard tool calls. This guide covers the documented path for adding and testing a custom internal server, with connection, authentication and network requirements included only where current Seahorse documentation confirms them.

How to Add a Custom Internal MCP Server to a Seahorse Cloud Agent

Key takeaways

  • Register uncovered servers: Seahorse Cloud documents custom MCP-server registration for servers not covered by its GitHub, Atlassian, PostgreSQL, and Slack templates. It supports Standard I/O for a local process and Streamable HTTP for a remote server. (Custom create)
  • Check exposed methods and calls: A tool’s exposed-method list is populated by Test or an automatic list_tools call. Recent tool-call history includes the status, duration, and any error message. (Tool detail)
  • Plan for connection risks: Upstream changes, driver versions, network configuration, or authentication configuration can result in connection failures, timeouts, or mismatched tool specifications. (Custom create)

Seahorse Cloud’s documentation describes a route for registering a custom MCP server as an agent tool, then checking the methods it exposes and reviewing recent calls. This guide covers the documented setup choices and checks. Registration alone does not establish that a particular server will work in every environment; the documented failure risks make testing the integration important.


What Seahorse Cloud documents for custom MCP servers

Seahorse Cloud documents direct registration of MCP servers not covered by its GitHub, Atlassian, PostgreSQL, and Slack service templates. Its custom-tool guide supports two transports: Standard I/O (stdio), which runs a local process, and Streamable HTTP, which connects to a remote server. (Custom create)

For an internal service, the relevant choice is whether the MCP server is a local process or a remote server. The guide warns that upstream changes, driver versions, network configuration, and authentication can lead to connection failures, timeouts, or mismatched tool specifications. It also states that Seahorse does not guarantee operation or compatibility for MCP servers outside its officially supported scope. (Custom create)

A registered tool’s detail page shows its configuration, exposed methods, connected agents, and call history. The exposed-method list is populated by Test or an automatic list_tools call; recent tool-call history is available for debugging and auditing. (Tool detail)


Choose a transport and prepare the connection

Standard I/O runs a local process and communicates through stdin and stdout. The guide lists internally developed MCP servers, internal tools, and servers distributed through npm or pip as examples. Streamable HTTP connects to a remote HTTP endpoint, including an MCP server provided as a service. (Custom create)

The setup guide recommends checking that the server responds before configuring it in Seahorse. For stdio, run the Command and Args in a local terminal to confirm that the MCP server process starts. For HTTP, send a request to the Endpoint URL with curl and confirm that it responds. These checks help identify basic server or endpoint issues; they do not replace checking the integration after registration. (Custom create)

For stdio, check whether the required executable is installed and accessible on the Seahorse agent server. The guide names npx, python, and uvx as examples. For HTTP, review whether the network configuration permits access, including relevant proxy, firewall, VPC, security-group, port, and domain settings. The guide also says to verify the TLS certificate and use Verify SSL: true in operating environments; Verify SSL: false is for development only. (Custom create)

Check the MCP server’s required credentials and permissions as part of preparation. The setup guide calls out issued credentials, such as API tokens, database accounts, or SSL certificates; approval of required scopes or roles by the external service; and token expiry and rotation. It warns that upstream changes, driver versions, network configuration, and authentication can cause connection failures, timeouts, or mismatched tool specifications. (Custom create)

If the server uses a package version, the guide recommends pinning it—for example, @pkg@1.2.3—rather than using @latest, which may break after an upstream change. (Custom create)


Register the server and configure its connection

The documented console path is Console → Agent Ops → Tool Management → + Create Tool, then select Custom under Quick Setup. The custom-tool guide lists Name, Transport, Description, and Access among the setup fields. Access has Private and Public options; the template guidance says Access defaults to Private and advises reviewing which agents in the tenant should use a tool before making it Public. (Custom create; Templates)

Choose the transport that matches the server: Standard I/O for a local process or Streamable HTTP for a remote server. The custom-tool guide marks the Transport field as fixed in its configuration example. (Custom create)

Standard I/O

For Standard I/O, configure the Command and Args used to run the local process. The guide’s examples of commands include python, node, npx, /usr/bin/python3, and uvx. It also documents two kinds of configuration values: User Input values are entered in the tool creation or editing screen; System Default values are set by an operator in Environment and automatically injected at runtime. (Custom create)

The custom-tool guide advises passing secrets through System Default environment variables rather than putting tokens or passwords in Args, because arguments are recorded in audit logs. (Custom create)

Streamable HTTP

For Streamable HTTP, the guide lists an Endpoint URL, a timeout in seconds, and a Verify SSL setting. It also documents adding request headers. Examples include Authorization: Bearer {token} and X-API-Key: {key}. For operating environments, the guide recommends SSL verification; it reserves disabling verification for development. (Custom create)

Seahorse’s template documentation says its templates accept and pass through values needed for connection and authentication; they do not create the supplier-side permissions, apps, or tokens. The template guide distinguishes User Input credentials, entered directly by the user, from System Default values, which are automatically used at runtime. These template details describe those integrations, not a general credential setup for every custom server. (Templates)

The custom-tool guide instructs administrators to click Save to register the tool in the tool list. (Custom create)


Attach the tool to an agent and test it

The Agent creation guide documents selecting multiple registered MCP and built-in tools. It also states that only inference endpoints with tool-call support can be connected to an agent. (Agent create)

To check what the server exposes, use Test on the registered tool. The tool-detail guide says Test performs an MCP handshake and automatically calls list_tools; when successful, the supported-method list is updated. Method details can include a description, an input schema, and required arguments. (Tool detail)

A populated method list shows which callable methods the MCP server exposes. It does not, by itself, establish that a method will produce the intended result in your target service. For an end-to-end check, the custom-tool guide recommends sending a query in Agent Chat that is likely to use the tool. Where the action has consequences, check the result in the target service as well as the agent’s response. (Custom create)

The tool-detail guide documents recent-call statuses of Success, Failed, and Timeout, along with duration in milliseconds and an error message when a call fails. It describes recent tool-call history as useful for debugging and auditing. Use these records to distinguish a successful call from a failed or timed-out one. (Tool detail)

The Agent creation guide documents tool selection and the tool-call-enabled endpoint requirement. The tool detail guide documents Test, the exposed-method list, and recent-call history.


Troubleshoot failures and set a safe operating boundary

If the server does not connect or a call fails, review the documented risks: upstream changes, driver versions, network configuration, and authentication configuration. These can result in connection failures, timeouts, or mismatched tool specifications. (Custom create)

Also check that the required executable is available to the Seahorse agent server, that the relevant network configuration permits access, and that required credentials have been issued and the necessary scopes or roles approved. Review credential expiry and rotation, and for HTTP check TLS certificate validity and the SSL verification setting. The guide identifies these as setup considerations; it does not present a universal diagnostic sequence for every failure. (Custom create)

If the exposed methods do not match expectations, review the method list populated by Test or list_tools. For failed or timed-out calls, consult the recent-call history and its documented status, duration, and error information. (Tool detail)

FAQ

Does Seahorse Cloud support custom internal MCP servers?
Yes. Seahorse documents direct registration for MCP servers not covered by its GitHub, Atlassian, PostgreSQL, and Slack templates. It supports Standard I/O for a local process and Streamable HTTP for a remote server. (Custom create)

How can I verify which tools the server exposes?
Run Test on the registered tool, or check the method list populated by an automatic list_tools call. The tool detail page also provides recent-call history, including Success, Failed, and Timeout statuses. (Tool detail)

What can cause a custom-server connection or tool-specification problem?
The setup guide identifies upstream changes, driver versions, network configuration, and authentication as possible causes of connection failures, timeouts, or mismatched tool specifications. It also states that operation or compatibility is not guaranteed for MCP servers outside Seahorse’s officially supported scope. (Custom create)

Which agent endpoints can use registered tools?
The Agent creation guide says only inference endpoints with tool-call support can be connected to an agent. It also documents selecting registered MCP and built-in tools. (Agent create)


지금 체험

See how these criteria hold up on the real thing.

지금 체험