n8n added MCP client support earlier in 2026, but what's changed across the 2.26.x and 2.27.x cycles is the depth of that integration — and the fact that it now runs in both directions. That's a meaningful shift for anyone building agent-driven automation, so it's worth slowing down on the actual release notes instead of a vague "AI-powered" headline.
MCP, briefly
If you missed our earlier write-up on MCP, the thirty-second version: Model Context Protocol is an open standard that solves the N×M integration problem. Before MCP, connecting AI tools to services meant a custom connector for every pair. With MCP, you build one interface per service and any MCP-compatible agent can call it — the USB-C analogy Anthropic uses is accurate enough to be useful.
n8n began shipping MCP compatibility in the 2.22.x range. The question is what 2.26.x and 2.27.x added on top.
What the 2.26.x stable releases added
n8n 2.26.0 introduced per-user connections to MCP servers from the registry. Previously, MCP server credentials were shared at the workspace level. Now each team member connects to MCP-compatible tools with their own credentials — cleaner permissions, cleaner audit trails if something goes wrong.
The same cycle expanded the MCP registry itself: servers can now extend existing n8n credentials rather than requiring separate setup, and team project context gets surfaced through MCP search. Sub-agent delegation was also formalized in this release — a workflow can hand off to a specialized sub-agent inline, with session linkage so execution context carries through. This is the kind of plumbing that makes MCP viable in shared environments rather than just developer sandboxes.
The current stable release is n8n 2.26.9, dated June 22, 2026.
What the 2.27.x beta adds on top
n8n 2.27.0 shipped as a pre-release on June 16, 2026. The current beta is 2.27.3 (June 19, 2026). Three additions are worth noting:
explore_node_resourcesMCP tool. An AI agent connected to your n8n instance can now ask "what integrations do you have available?" and get a structured answer. Agents discover capabilities dynamically rather than relying on hard-coded node names.- Tags on MCP servers. Once you're running more than a handful of connected services, you need to route agent requests to the right tool. Tags make MCP server registrations filterable and searchable.
- OAuth2 authentication for the MCP Server Trigger. When you expose an n8n workflow as an MCP server endpoint, external AI agents authenticate via OAuth2 rather than shared API keys. That's the security posture you need in production.
These features are in beta as of today. If you're on the stable 2.26.9 branch, they'll land in the next stable point release.
The two-way picture
n8n can now be an MCP client — your n8n agents consume tools from external MCP servers — and an MCP server — external AI agents call your n8n workflows. Most platforms have picked one side.
The community tooling around this already shows the scope. The open-source n8n-mcp project exposes 1,845 n8n nodes via MCP — 816 core nodes plus 1,029 community integrations — along with 2,352 pre-built workflow templates and 265 AI-capable tool variants. A separate workflow builder MCP server provides 15 workflow management tools so AI assistants can create, activate, monitor, and audit n8n automations through natural language.
Practically: a Claude agent could query which n8n integrations are available via explore_node_resources, locate the Stripe node, assemble a refund workflow for a flagged transaction, run it, and log the result — without a developer touching the configuration. The MCP Server Trigger with OAuth2 means the agent authenticates cleanly, not with a shared key buried in a config file.
Does this change the platform decision?
Zapier is still faster to configure for two-step automations. Make is still more visual for complex branching. But neither ships native bidirectional MCP support at this depth.
We covered the full n8n vs Zapier vs Make decision in an earlier post. The MCP additions don't change that framework — they amplify the "n8n for agent-heavy workflows" conclusion that was already there. If your roadmap involves LLM-powered agents that need to reason about and route across your tool stack, the ability to expose that stack as MCP-callable endpoints is significant.
One thing that hasn't changed: the self-hosted Community Edition remains free with no execution caps, and the MCP features don't require a cloud subscription. If you're comparing per-seat costs, that still matters.
What to actually do with this
If you're already running n8n on 2.26.x, the MCP registry features are live. The useful next question is which of your workflows are candidates to expose as MCP tools — the answer is usually anything your team currently triggers manually on a predictable schedule or condition. Repetitive, context-heavy tasks where a human currently decides "yes, run this now" are the best starting point.
If you're evaluating n8n for the first time, the bidirectional MCP support is a concrete differentiator to weigh in the platform comparison, not a marketing claim. You can verify it in the official GitHub release notes.
If you want an outside read on whether your current stack is worth building on or whether the underlying automation plumbing needs rethinking before adding agents, drop us a line. That's the kind of audit we do before recommending any direction.
— Cole