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

Long-running MCP tools (like ResearchClaw’s 23-stage pipeline) should return immediately with a run_id rather than blocking — the caller polls status via a separate get_status tool. Expose gate-approval as dedicated tools (approve_gate, reject_gate) so the orchestrating agent can act as a programmatic HITL reviewer. Expose sub-capabilities (citation verification, claim verification, literature search) as standalone tools callable outside the pipeline to maximize reuse.