LoomCycle MCP Server Tool icon

LoomCycle MCP Server Tool

Registers an MCP server in the loomcycle substrate (idempotent), then exposes a tool that delegates to a loomcycle agent using that MCP server

LoomCycle MCP Server Tool

AI-generated

Summary

Register or reuse an HTTP-based MCP server in LoomCycle, then expose an AI tool that sends the parent agent’s prompt to a configured LoomCycle agent with access to that server’s tools. The delegated prompt is treated as untrusted, and the tool returns the agent’s final text.

Inputs

  • Tool Name (required) — Name shown to the parent AI Agent when it chooses a tool.
  • Tool Description — Description shown to the parent AI Agent to help it decide when to call the tool.
  • MCP Server Name (required) — Unique LoomCycle registration name. The delegated agent is allowed tools matching mcp__<name>__*. If a registration with this name already exists, it is reused; otherwise the node creates and promotes one.
  • Transport (required) — MCP transport: HTTP or Streamable HTTP (default). Stdio is unsupported for dynamic registration and must be configured in loomcycle.yaml.
  • URL (required) — MCP server endpoint URL. Its hostname must be permitted by LoomCycle’s HTTP host allowlist when creating a registration.
  • Headers — Optional headers attached to MCP calls. Values support LoomCycle environment-variable and per-run substitutions; environment-variable names found in header values are logged for LoomCycle configuration.
  • Loomcycle Agent (required) — Name of the LoomCycle agent spawned for each tool call.
  • Cleanup On Workflow End — When enabled, attempts to retire the MCP registration when the workflow execution ends. Defaults to false, leaving the registration in place.
  • prompt (required) — The task text passed by the parent AI Agent to the delegated LoomCycle agent.

Output shape

An AI Tool connection for an AI Agent; direct execution returns one item per input item with the delegated agent’s final text in json.result.

The node registers or reuses the MCP server before exposing or executing the tool. Direct execution reports per-item failures in json.error. The agent run is streamed and drained to completion; stream error events become errors. If cleanup is enabled, retirement errors are ignored.

Links

Discussion