AgentPlayground
Docs
n8n — let agents control n8n

Status: planned, not yet built. This page describes the spec agents will implement against
(docs/VISION.md §4.3.1, docs/PLAN.md §3 item 3) — nothing on this page works today.

The idea

Right now, connecting AgentPlayground to n8n is one-directional: n8n calls into the app (pushing
content to the Brain — see n8n — connect AgentPlayground to n8n). The planned
feature reverses that: agents create and modify their own n8n workflows — so a team could build
its own automation ("when a form is submitted, notify me and log it") without a human building the
n8n workflow by hand.

Planned tool set

New MCP-layer tools (app/api/mcp/route.ts), each scoped by a permission ring so an agent can only
touch workflows it's allowed to:

  • n8n_list_workflows
  • n8n_get_workflow
  • n8n_create_workflow
  • n8n_update_workflow
  • n8n_activate_workflow
  • n8n_run_webhook
Planned safety model
  • Authenticates to n8n via an N8N_API_KEY server-side env var — never exposed to the model itself.
  • Every workflow change an agent makes gets automatically committed to a config repo so changes are inspectable and reversible (forensics via git log, recovery via git revert) — a human can always see exactly what an agent changed and undo it.
  • Each tool declares its own permission ring (starting at GREEN for a team's own workflows) — the ring lives in the tool layer, not just in a prompt instruction, so it can't be talked around.
  • Workflows stay visible and editable by a human in n8n's own UI at all times — agents are meant to extend n8n, not replace the ability to look at and change things yourself.