Best AI Agent Skills for n8n Workflow Automation (2026)

Ask Claude to build you an n8n workflow without giving it any node knowledge and you get JSON that looks right and imports broken. The parameter names are plausible. The node types are almost correct. Nothing runs.

That happens because n8n is large. Its own MCP server's README counts 2,541 nodes: 832 core and 1,709 community. No model has that schema memorised, and n8n's error messages when a parameter is wrong are not the kind that point you at the fix.

So the useful entries in this category are not automation skills in the abstract sense. They are schema delivery. They give the model the node catalogue, the parameter shapes, and a way to validate a workflow before it touches your instance.

Four layers, and only two of them are skills

Directories collapse all four into one list. That is how people end up trying to npx a self-hosted platform. Our own database does it too, and I flag the specific cases below.

Node knowledge: the layer that fixes the problem

n8n-MCP

czlonkowski/n8n-mcp, MIT, 22,751 stars, last pushed 19 August 2026. This is the entry worth your attention and the one to install first.

It exposes n8n's node documentation, properties and operations to the model as callable tools, so Claude looks up the Slack node's actual parameter schema instead of guessing at it. The project claims 6,524 passing tests, which for a schema-delivery server is the number that matters, since the failure mode is silent drift from n8n releases rather than a crash.

Install route, verified against the npm registry today: the package is n8n-mcp, currently v2.73.0, published from czlonkowski/n8n-mcp, and it ships a stdio binary.

claude mcp add n8n-mcp -- npx -y n8n-mcp

Our database lists this as npx -y czlonkowski/n8n-mcp, substituting the GitHub path where npm wants a package name. That string fails. The one above works.

There is also a hosted version at dashboard.n8n-mcp.com if you want to try it before self-hosting, and a self-hosting guide covering npx, Docker and Railway. If your n8n sits behind Cloudflare Access, the server takes a service token and sends it as CF-Access-Client-Id and CF-Access-Client-Secret headers, scoped to the n8n API origin so a split webhook host does not receive it.

Deep dive: /skills/n8n-mcp

n8n-as-code

EtienneLescot/n8n-as-code, MIT, 1,505 stars, pushed 21 August 2026. A different bet: put workflows in your repository as TypeScript, and treat the n8n instance as a deploy target rather than the source of truth.

The skill half covers 537 nodes with full schemas. The CLI half is where the value sits, because it handles the thing visual automation is worst at, which is having more than one environment.

npx --yes n8nac env add Dev --base-url https://n8n.example.com --workflows-path workflows/dev
npx --yes n8nac pull <workflow-id>
npx --yes n8nac push workflows/dev/my-workflow.workflow.ts --verify
npx --yes n8nac promote --from Dev --to Prod --dry-run

Inside Claude Code it installs through its own marketplace with /plugin install n8n-as-code@n8nac-marketplace. There is a VS Code extension too. The API key goes in over stdin rather than as an argument, which is a small detail that tells you the author has thought about shell history.

Reach for this when workflows are production infrastructure that needs review and rollback. If you are prototyping a single automation, it is more ceremony than you want.

Deep dive: /skills/n8n-as-code

The three that look the same and are not

Three more n8n MCP servers sit in our database, and star count is a bad way to choose between them. Maintenance date is better.

A stale MCP server in this category is worse than no server, because it hands the model confidently wrong schemas for nodes that have since changed. That is harder to debug than an empty context, where at least the model admits it is guessing.

All three are unreviewed community tier. None of them is a reason to skip n8n-MCP.

Platforms, which are not installable as skills

n8n

n8n-io/n8n, 201,535 stars, one of the largest repositories in our whole directory. Fair-code under the Sustainable Use Licence, which is not OSI-approved, so read it before you build a product on top.

Our database gives it an npx -y n8n-io/n8n install string. n8n is an application you self-host or buy, not a server you register with Claude. npx n8n will start the app on your machine; it will not give Claude anything. Use n8n-MCP for that.

Deep dive: /skills/n8n

Activepieces

activepieces/activepieces, 23,966 stars, pushed 21 August 2026. Around 400 MCP servers usable as agent tools, and a workflow builder aimed at the same job n8n does. There is no activepieces package on npm, so the install string in our database returns a 404. Self-host it with Docker or use their cloud.

Worth considering if you want the automation platform and the MCP layer from one vendor rather than bolted together.

Deep dive: /skills/activepieces

OpenOps

openops-cloud/openops, 1,059 stars, active. No-code FinOps automation: cloud cost workflows rather than general glue. Narrow by design, and the narrowness is the point if cost is what you are automating.

Deep dive: /skills/openops

When visual workflows stop being the right shape

Conductor

conductor-oss/conductor, Apache-2.0, 32,104 stars, actively maintained, Java. A workflow orchestration engine in the durable-execution family, with retries, versioned definitions and state that survives a restart.

The honest framing: this is not an n8n competitor for most people. It is where you land when a visual workflow has grown into a distributed system with failure semantics you need to reason about, and dragging boxes has stopped helping. Apache-2.0 and a real governance model, which matters more here than in the categories above.

Deep dive: /skills/conductor

Credentials are the part nobody plans for

Every entry on this page eventually holds an n8n API key, and an n8n API key is not read-only. It can create workflows, and workflows execute code.

Three habits worth adopting before you install any of these. Give each server its own key so revoking one does not break the rest. Point them at a staging instance until you trust the generated workflows. And check what a workflow does before activating it, because an agent that writes automation is an agent that writes automation that runs without you.

Related reading: agent skills for multi-agent orchestration, our n8n and Zapier platform comparison, and the full list for developers.

The order I would go in

• Install n8n-MCP with the corrected npm command. On its own it fixes the invented-parameter problem, which is most of the pain.
• Point it at a staging n8n instance, not production, until you have watched a few generated workflows run.
• Add n8n-as-code once workflows are worth reviewing in pull requests. Skip it before then.
• Ignore the three older MCP servers unless one of them does something specific that n8n-MCP does not.
• Consider Conductor only when you have hit a limit you can name out loud, not because it sounds more serious.

FAQ

Which n8n MCP server should I install?

czlonkowski/n8n-mcp. It is the most maintained, the most tested, and the only one in this category with a hosted option for evaluation. The others are either stale or narrower.

Why does the install command in your database not work?

Because the ingestion pipeline builds install strings from the GitHub owner and repo, and npm wants a package name. For n8n-MCP those differ, so the generated string fails. We are correcting these entries. Until then the repository README is the authority, and this page carries the verified command.

Is n8n itself a skill I can install?

No. n8n is a self-hosted or cloud application. A skill or MCP server is what lets Claude talk to it. Install n8n normally, then install n8n-MCP so the model can read the node catalogue.

Do I need n8n-as-code if I already have n8n-MCP?

Only if you want workflows in version control. They solve different problems: one teaches the model what the nodes do, the other treats workflows as reviewable source with environment promotion.

How risky is giving an agent an n8n API key?

More than it looks. That key can create and activate workflows, and n8n workflows can run arbitrary code nodes and reach your internal network. Treat it like a deployment credential rather than a read token: separate key per server, staging first, and read what got generated before switching it on.