Automatic discovery and refresh
- Orca Engine registers agents that expose the required
root_agent
entry point and marks them as shareable through MCP. - The managed server regularly discovers eligible agents and publishes them as MCP tools. Default tool names follow the pattern
agent_<tenant>_<namespace>_<agent>
when you do not supply overrides. - Server-Sent Events (SSE) signal catalog changes. When the MCP server reports
tools changed
, Orca Engine reloads cached metadata so clients immediately see the latest instructions and tool hints.
Metadata surfaced to MCP clients
- Each exported tool includes the agent name, description, and supported session mode so callers understand the conversational context before invoking it.
- When an agent declares structured input or output channels, the Remote MCP Server maps that contract into MCP tool parameters so clients know which fields to supply.
- Keep prompt instructions concise and results deterministic where possible so downstream tools can parse responses reliably.
Customize tool expose metadata
Provide descriptive names and summaries to make the MCP tool easier to be used by LLM agents.- Set
MCP_TOOL_NAME
andMCP_TOOL_DESCRIPTION
as environment variables in the agent configuration through CLI deployment flags. - When present, the Remote MCP Server uses these values instead of the default derived name and description.
- Administrators can require both variables before publishing an agent so only well-documented tools appear in Remote StreamNative MCP server.
Prepare an agent for MCP publication
- Follow the Google ADK agents guide or OpenAI agents guide to produce a deployable artifact.
- Store model credentials as StreamNative Cloud secrets and bind them with the
--secrets
flag during deployment. - Add
MCP_TOOL_NAME
andMCP_TOOL_DESCRIPTION
to the agent environment, then redeploy so the discovery cycle can pick up the customized metadata.
Control discovery with MCP features
- The Remote MCP Server exports shareable Orca agents by default when your identity can reach the associated tenant and namespace.
- To skip agent discovery from a client, send the
X-MCP-Features
header during connection setup and omitagents-as-tools
from the list. - Provide the feature explicitly—
X-MCP-Features: pulsar-admin,agents-as-tools
—when you want to mix agent tools with other MCP capabilities while keeping control over optional features.
Consume the agent from MCP clients
- MCP clients invoke the tool using standard
call_tool
requests. The managed server forwards payloads to the corresponding Orca agent and streams responses back. - The MCP server enforces StreamNative Cloud permissions, so users only see agents they are authorized to access.
- Use StreamNative observability dashboards and
snctl agents status
to monitor traffic routed through MCP.