async-mcp-pattern-for-long-running-pipelines

Long-running MCP tools (like a 23-stage research pipeline) should return a run_id immediately rather than blocking. Separate status, get_output, resume, approve_gate, and reject_gate tools allow ORACLE Hermes to poll progress and act as HITL reviewer programmatically. All responses return structured JSON, not raw stdout. This pattern is now the canonical design for any pipeline-style MCP server.

Related

• autoresearchclaw-mcp-server-integration-pattern
• researchclaw-mcp-server-11-tools-interface-design
• oracle
• async-mcp-tool-pattern-for-long-running-pipelines
• autoresearchclaw-mcp-server-11-tools-interface-design
• autoresearchclaw-gate-stages-hitl-mcp-pattern
• mcp-async-pipeline-must-return-run-id-not-block