14 capability Skills and one router, working through instance-level MCP.
Official n8n Skills
From Draft to Execution Evidence
Connect workflow design, node configuration, testing and diagnosis into a process you can inspect. Start with a data transformation, an existing failed execution or a dedicated test ledger in an n8n instance you can access.
Hosts listed by the author
- Claude Code
- Codex
- OpenCode
- GitHub stars
- 561
- Version
- 1.3.0
- Checked
- Categories
- EngineeringProduct & Project
Start with an unpublished workflow you can inspect
The first task connects no external services. It checks whether three inputs reach the expected branches while exercising connection, creation, saved-state inspection and testing. Add real business actions after each step has evidence.
Pick a task and get started
Install the plugin inside Claude Code, configure and authorize the instance, then choose a task. Replace bracketed project and execution details with your own values; placeholders are not real resources.
Start with Manual Trigger, If and Edit Fields. This saves an unpublished workflow in your instance and runs its data transformations, without credentials for email, databases or model nodes.
Run each installation line inside Claude Code
Run each line in a Claude Code session, then enter the instance URL and complete authorization below. Adding the marketplace does not install the plugin. Generic npx skills add does not supply the same hooks and connection configuration.
Read the author’s setup guide/plugin marketplace add n8n-io/skills /plugin install n8n-skills@n8n-ioConnect the instance and check permissions
When prompted, enter the n8n base URL, such as https://your-team.app.n8n.cloud, without a path or trailing slash. The plugin appends /mcp-server/http. An instance owner/admin first enables Settings → Instance-level MCP. In Claude Code, run /reload-plugins, then open /mcp, select n8n-mcp and Authenticate. Complete browser sign-in and grant the capabilities needed for your task. Verify a read-only query before checking creation, validation and testing tools. Existing workflows also need Available in MCP enabled for full reading and operations; finding their names in search does not establish that access.
Read n8n’s connection and permission guideGive your AI this first task
These natural-language tasks invoke installed Skills. The assistant calls the tools over MCP; tool names are not terminal commands.
Use using-n8n-skills-official and load n8n-workflow-lifecycle-official, n8n-node-configuration-official and n8n-expressions-official. Create an unpublished workflow named “Validate sample orders” in my personal project. First confirm the connected n8n instance and the available editing, validation and testing tools. Report missing tools; do not substitute execute_workflow for a test. Check that orderId is a nonempty string, amount is a number greater than zero, and currency is a three-letter string. Accepted items should return orderId, amount, uppercase currency and status="accepted". Rejected items should return orderId (null when missing), status="rejected" and a specific reason. Use only Manual Trigger, If, Edit Fields and pure expressions, processing each item separately. Do not add external services, Data Tables, files, commands or sub-workflows. Read get_workflow_sdk_reference, then use search_nodes and get_node_types for this instance's node definitions. Validate the complete SDK code before saving. Read back get_workflow_details to verify the saved nodes, connections, both branches and output fields. Run three separate test inputs: {"orderId":"demo-001","amount":1200,"currency":"usd"}; {"orderId":"demo-002","amount":-5,"currency":"usd"}; {"amount":25,"currency":"eur"}. Use prepare_workflow_pin_data for the expected shape, generate pinData keyed by the actual trigger name with each item wrapped in {"json": ...}, then call test_workflow. The expected outcomes are acceptance, rejection for the amount, and rejection for the missing order ID. Fetch get_workflow_execution with includeData=true and compare the actual branch output with each expectation. After a fix, repeat the same inputs. Deliver the workflow ID, URL, project, unpublished status, validation warnings, execution IDs and output comparisons. State which nodes were pinned and which ran. Do not publish or activate.
Check what you receive
- An unpublished workflow in the instance, with its ID, URL, project and a check of the saved nodes and connections.
- Execution records comparing all three inputs with expected results. Keep failures and unverified items visible instead of reporting only a validation pass.
More about this skill
n8n-io/skills is n8n’s official collection of operating guidance. The reviewed commit contains 14 capability Skills plus using-n8n-skills-official, which routes work to the relevant guidance. It covers workflow building and maintenance as well as a separate set of instructions for n8n Agents. Installing the files teaches an assistant what to inspect and in what order; saving and running workflows requires a connected n8n instance.
The central method is to query the instance before generating configuration. The assistant reads the Workflow SDK reference, searches for nodes, retrieves parameters for the selected operation, and then writes and validates the workflow. Node definitions change between versions. An old example cannot replace the live definition, and resources such as tables, channels and documents need real IDs rather than plausible names.
Delivery involves several kinds of evidence. Code validation checks structure. Reading a saved workflow confirms that nodes and connections reached the instance. Sample execution shows the branch data actually took. Incorrect business rules, missing paths and fields replaced by an upstream node can survive a schema check. This page therefore asks for execution records and output comparisons as well as the saved workflow.
Testing and going live are distinct operations. test_workflow substitutes supplied outputs for triggers, credentialed nodes and HTTP Request nodes; other nodes generally execute. Data Table writes, sub-workflows, files and commands deserve an explicit check. The manual mode of execute_workflow also runs real logic. Publishing creates an active version for production execution. An unpublished workflow can still run, and passing a pinned test does not verify the real external service.
The native Claude Code plugin adds guidance at session start and around relevant tool calls. These reminders help surface node configuration, testing and remaining risks; their behavior still depends on the host and available tools. A generic skills installation supplies the files without the same hooks, and it does not enable MCP or complete authorization for you.
The first-use tasks here focus on workflow capabilities: a pure transformation, diagnosis from saved evidence, and a queryable test ledger. An AI Agent node inside a workflow and a standalone n8n Agent are different objects with different tools and configuration. Enter the corresponding Skill when you need models, tool permissions or channels, rather than treating an ordinary workflow test as proof that a complete Agent is ready for production.
Sources and verification
Sources checked
Checked plugin 1.3.0, the inventory of 15 SKILL files, the Skills and references used here, and n8n’s official connection and tool documentation against the recorded source snapshot.