n8n Agents is a skill for Claude that goes deep on the LangChain nodes in n8n. It covers the choice between an AI Agent, a Basic LLM Chain, a Text Classifier and an Information Extractor, and the model, memory, tool and output parser slots of an agent.
It also covers tool names and descriptions as part of the prompt, the parameters an agent fills with $fromAI, structured output with autoFix, memory keys, RAG, human approval nodes, chat bot topologies and persisted agents managed through the n8n MCP tools.
What You Get
- A decision table for AI Agent, Basic LLM Chain, Text Classifier, Information Extractor, Sentiment Analysis and Summarization Chain
- Rules for tool names and descriptions, and for the $fromAI parameter name, description and type
- Defaults for structured output with a manual JSON Schema, autoFix and a separate fixer model
- Memory guidance covering the buffer window, session keys and when to use Postgres or Redis chat memory
- Reference files for tools, sub-workflow tools, system prompts, memory, human review, chat bots, RAG and examples
- A pre-ship checklist and an anti-pattern table
Common Use Cases
- Building a Slack or Discord support bot as an n8n workflow with a bot-user filter to stop reply loops
- Routing incoming text to several branches with one Text Classifier node
- Wrapping refunds or email sends in a human approval step that shows the real tool parameters
- Fixing an agent that skips a tool or hits the iteration cap
- Creating and publishing a persisted n8n agent through the n8n MCP tools
What Makes It Unique
It is tied to n8n’s own node types and connection names (ai_languageModel, ai_memory, ai_tool, ai_outputParser) and to the n8n-mcp tools that build and validate them. It states opinions a general prompt guide would not: send routing to a Text Classifier, keep per-tool advice in the tool description, and show approvers the literal tool parameters. It points to sibling skills for code tools, binary data and sub-workflows.
Under The Hood
The skill is instructions plus reference markdown files, with no scripts of its own. It tells Claude to call the n8n-mcp tools, which search nodes, validate configuration, patch workflows, run test executions and read saved workflows from your n8n instance. Those calls reach your n8n server, and the persisted-agent section can create, run, validate and publish agents with real credentials, so a live call runs real tools. It tells Claude to publish only when you ask.