n8n is the automation tool engineers reach for when they want to build AI agents on a canvas but still drop to code, and run the whole pipeline on their own infrastructure.
Can n8n build AI agents without code?
Yes, n8n builds production AI agents through its LangChain-based AI Agent node, which orchestrates a chat-model sub-node, a memory sub-node, and tool sub-nodes as one workflow execution, reasoning and calling tools without code, while still letting you drop to code when the visual abstraction runs out. To build an AI agent in n8n, you add the AI Agent node, attach a chat model, attach memory, and attach tools.
The escape hatch is the differentiator. A Code node running JavaScript or Python, or a raw HTTP Request node, means a missing connector never blocks you. The model layer is provider-agnostic: you swap OpenAI with GPT-4o, Anthropic Claude, Google Gemini, or local Ollama by changing one sub-node, with no rewrite of the surrounding workflow, and the same applies to Groq, Mistral, Azure OpenAI, and AWS Bedrock.
How does the AI stack fit together (RAG, memory, tools)?
n8n assembles a full RAG stack natively, where embeddings nodes, document-loader nodes, and text-splitter nodes such as the Recursive Character Text Splitter handle ingestion, and native vector-store nodes handle retrieval, so the AI Agent node can ground answers in your own data without external glue. The supported vector stores include Pinecone, Qdrant, Supabase, PGVector, Redis, Weaviate, Chroma, Zep, Milvus, and MongoDB Atlas, plus an in-memory store.
Tools are just integrations. Any of n8n's 400+ built-in integrations can become an agent tool, as can a sub-workflow via the Workflow Tool or arbitrary logic via the Code Tool, so agents take real actions rather than only describing them; community nodes raise the effective count further but are not part of the official figure. Memory is more limited: the built-in memory nodes, including the Simple window buffer, Postgres Chat Memory, and Redis Chat Memory, are per-session and ephemeral. There is no managed long-term agent memory, so durable cross-run state is do-it-yourself via Postgres, Redis, or a vector store.
Is n8n free if you self-host it, and what does Cloud cost?
Yes, n8n's self-hosted Community Edition is free under its fair-code Sustainable Use License, with unlimited workflows, steps, and executions, so you pay only for the infrastructure you run it on; it is free to run, but it is not MIT or OSI open source. The official Self-hosted AI Starter Kit, a Docker Compose bundle of n8n plus Ollama plus Qdrant, is the fastest way to stand up a fully local AI pipeline.
n8n Cloud bills per successful workflow execution: one full run counts as one execution regardless of node count or data volume, and failed or manual test runs do not count, across tiered Starter, Pro, Business, and Enterprise plans. The permanent Cloud free tier was removed in late 2025, though a time-limited trial remains. On Cloud, the AI Workflow Builder is metered in AI credits that vary by tier, roughly 50 on the entry tier up to around 1,000 at the top, per n8n's own pricing; the Builder is also usable self-hosted with your own API key, and LLM token cost is always separate, paid to the model provider or absorbed by self-hosting Ollama.
Does n8n work with ChatGPT, Claude, and Gemini (and MCP)?
Yes, n8n connects OpenAI with GPT-4o, Anthropic Claude, and Google Gemini as swappable chat-model sub-nodes, alongside local Ollama, Groq, Mistral, Azure OpenAI, and AWS Bedrock, so you change the model you run by swapping a single sub-node rather than rewriting the surrounding workflow. In our builds we keep the model layer swappable so a provider price or quality change is a one-node edit.
MCP support is bidirectional and first-class: the MCP Server Trigger exposes workflows as tools to external clients such as Claude Desktop, ChatGPT, and Cursor, while the MCP Client Tool lets agents call external MCP servers. A native instance-level MCP server entered public preview in late April 2026, available on all editions including self-hosted Community on recent Community versions, and can build, validate, test, and publish workflows from natural language. Human-in-the-loop gating is built in: any agent tool can be marked to require approval, pausing the workflow until a person approves or rejects high-impact actions. The three platforms diverge cleanly on the axes that decide a build: on billing, n8n charges per execution, Zapier charges per task, and Make charges per credit; on hosting, n8n is self-hostable, Zapier is cloud-only, and Make is cloud-only with an Enterprise on-prem connectivity agent; on build surface, n8n is a canvas plus a code escape hatch, Zapier is linear no-code, and Make is a visual scenario canvas.
Where does n8n end and a studio begin?
n8n gives you the canvas and the escape hatch, but it does not give you durable agent memory, deep agent-loop tracing, or automatic horizontal scaling, since queue mode with Redis brokers and workers is operational work you run yourself, and Source Control is Business and Enterprise only. Source Control is environment sync rather than Git version control with diffs and pull-request review, so that production gap, covering durable state, observability, scaling, and human-in-the-loop gating on destructive tools, is the AI automation and AI engineering work a studio owns, and if you are weighing whether to operate that path yourself, it is worth scoping a build before you commit.