Port Workflows Agent and MCP Governance Guide

On this page

Quick answer

Port’s AI action node documentation distinguishes AI and AI-agent modes, with explicit tools, optional MCP connectors, provider and model selection, and optional JSON Schema output. Govern each of those choices separately.

Publishing a workflow as an agent tool does not authorize every agent or user to trigger it. Keep workflow permissions, tool authority, catalog ownership, approvers, and downstream credentials narrow.

Control layers

  • Pin provider, model, prompt, input source, and data classification.
  • Allowlist tools and MCP servers by exact identity, transport, action, and resource scope.
  • Treat structured output as untrusted until schema validation passes; define failure on invalid output.
  • Map the triggering user or agent to workflow permission and catalog context.
  • Require human approval before consequential writes, deletion, deployment, spending, or external communication.
  • Log inputs, tool calls, outputs, approvals, errors, retries, and final state without exposing secrets.

Test a prompt-injected catalog field, unapproved MCP tool, wrong tenant, invalid JSON, tool timeout, duplicate resume, lost approver, and partial downstream write. A valid schema does not prove a fact or authorize the action.

Use the approval guide and readiness checklist.

Official sources

Source check: August 23, 2026. Verify current providers, models, modes, MCP connectors, schemas, permissions, and audit behavior.