MCP Integration Reference

This section is a formal reference for downstream MCP server implementers: what reaches your server, documented headers and request shape, and how to structure multi-server apps. It is not a walkthrough of the Agentsyx Creator UI.

For conceptual background, see What is MCP?.

In this reference

• Request path & trust boundaries — Who calls whom; what you must not re-implement.
• Inputs to downstream servers — Messages, arguments, headers, correlation.
• Injected parameters & metadata — Why context is auto-injected; always-on headers (and MCP mirrored args); optional Creator Parameter Tagging.
• Anonymous & unauthenticated users — Behavior when identity is limited or absent.
• ChatVault user merge & loadMyChats patterns — Trusted canonical id headers, idempotent merge, Upstash + worker (saveChat-style queue), read scope.
• MCP server metadata — tools/list, resources/list, optional prompts/list, tools-only vs widget servers, stability.
• MCP App (widgets) metadata — resources/read, _meta linking tools to URIs, embedding and CSP, callServerTool, host logging.
• Multi-connector architecture — Split UI vs backend; upstream tool mapping; Zapier/Make/n8n as separate MCP backends.
• Integration platforms — Zapier, n8n, Make patterns.
• Versioning & compatibility — Breaking vs safe changes; tools, resources, _meta; deprecation and multi-connector coordination.
• Troubleshooting — Common failure modes and what to log.

Implementation details are grounded in monorepo/apps/saas-starter; if this doc and code disagree, trust the code and open an issue on the docs repo.