N8n is all the rage now. Teenagers are creating workflows that automate whole departments and then sell them for thousands of dollars on Twitter. Managers who never wrote a line of code in their life automating workflows of their entire department overnight.
That sounds almost too good to be true. And it is, most of that is the ICO crowd, pivoted to another area of semi-honest appropriation of other people’s money.
N8N is a powerful tool, don’t get me wrong. It is just not all-powerful and definitely not a silver bullet. At least not yet.
Let’s look at what it really can help with and what things are better accommodated in other ways.
I’m going to start with an obvious observation. To automate a workflow you have to have a workflow. Seems redundant to say that but it is not. Majority of small and some medium size organizations don’t have clearly defined processes and therefore any attempt to automate those will fail no matter what tools are used.
The other side of this coin that is more common in large organizations is when there are processes but they are so complex and entangled or fuzzy that no one person can explain or understand them and thus again any naive attempts to automate are doomed to fail.
One might think — no problem, if there are no processes then I’ll create workflows in whatever tool and make people in the organization use that! Well, that approach is akin to putting the cart before the horse. I have to say since the cart in question is somewhat intelligent it may work for simple scenarios but unlikely to work for anything complex.
So that leaves us a field of organizations with sufficiently well defined digital processes of simple to moderate complexity. It’s not a small subset though.
Now let’s look at N8N and what it brings to the table. N8N is an open-source workflow automation tool. In simple terms, it lets you connect different apps, services, and data sources together without having to write a lot of code. You can build workflows where one event (like receiving an email, getting new data in a spreadsheet, or a webhook call) automatically triggers actions in other tools (like sending a message in Slack, updating a database, or starting an API call).
This description makes it similar to Zapier and Make — and they are similar. There are two key distinctions:
1. N8N is more flexible and technical. It allows for much more fine-grained control and thus for more complex flows to be automated. This naturally comes with a more steep learning curve.
2. Open source. That’s right — you can deploy it on your server and not pay anyone to run it. And if it lacks any functionality you might need — you can add it yourself. Technically it is an N8N Fair Code License, not OSI-approved open source so limitations apply.
So by its nature it sits between software engineers writing custom code and non-technical users creating workflows in a visual editor.
For the latter N8N is just one more option to choose from so I’m not going to delve into this. For software engineers N8N represents a nice alternative to coding as it allows a level of granularity in automation, debugging and logging similar to coding.
There are also workflow orchestrators that are code-first but also offer visual editors that make them competitors of N8N, but they are much heavier in terms of infrastructure and learning curve. Those are mostly used in the corporate world, for example Temporal/Union.ai and similar platforms. They are mostly used for durable workflows for complex, long running and mission critical orchestration.
The basics
Everything in N8N is a node, which represents a step in a workflow.
For example there are:
- trigger nodes: Start when a new email arrives
- action node: Send message to Slack
- function node: Run custom JavaScript
Workflow starts when the trigger node is triggered. It can be an event of some sort or you can programmatically start it using a webhook trigger. Webhooks have practical constraints in that providers often enforce response timeouts, so for long-running flows it’s better returning early and finishing work asynchronously.
Data moves through nodes step by step. Each node can change or enrich the data before passing it on.
Each node output can be frozen at any point (pinned) — so it always returns the same output, which is very convenient for development purposes when it is expensive or long to wait for a node each time.
Outputs of each node of each workflow run can be inspected — which is super convenient for debugging, as you can imagine. So convenient in fact that this is much better than code.
You can re-run a workflow starting from an arbitrary node. This is very useful when testing or debugging a flow as well and something that many code automations do not have. This typically requires pinned data or prior execution data so you can’t always arbitrarily start from any node with fresh upstream context but it is still a very nice feature.
Two useful building blocks here are “Execute Workflow” and “Split in Batches”.
“Execute Workflow” lets you call another workflow as a subroutine, pass it data, and get back its result. This promotes reuse and keeps large flows manageable.
“Split in Batches” processes large item arrays in chunks (e.g., 50 or 100 at a time) to avoid timeouts and rate limits and to control memory usage.
Here is an example:
As you can see N8N has pre-built nodes for most common services and needs. You don’t need to create integrations with Gmail, Slack, Google Docs or most CRMs, which, again, you don’t get in code automations.
For things that are not there yet you have http request node that allows you to call any kind of api. It handles authentication, headers and any sort of parameters that might be needed. There are some limits, like complex auth flows or streaming are not supported as well as websockets and very large payloads/downloads can hit memory limits, so for that you might have to create custom code.
And you can do this — there is a code node that allows you to add whatever you can’t achieve with prebuilt nodes.
AI capabilities
Since n8n became popular in the AI era — there is a huge AI component to it.
It comes in two flavors:
No-code nodes
There are lots of them for any type of LLM, vector database or AI service provider like HuggingFace.
For example a typical podcast pipeline would look like:
Audio Upload -> Whisper (transcribe) -> OpenAI (summarize) -> Notion (save notes)
There is a special type of node for agentic AI called AI Agent:
As you can see it has a model that can be set up to any LLM.
The main thing that determines the Agent behavior is its prompt.
Additionally it can have memory so that the agent remembers the context between calls, and a set of tools the agent can use. There are a lot of pre-built tools already that make for some very rapid prototyping. And of course you can stack several agents if you think one is not enough.
If the built-in functionality is not enough — you can leverage custom code nodes.
LangChain nodes
This type of node allows you to use LangChain’s chains and agents inside an N8N workflow. You will have to write code to use that but then you have the ultimate flexibility of coding.
Reliability and error handling
Many nodes support Continue On Fail so the flow proceeds and you can branch on errors. You can use an Error Workflow via the Error Trigger to catch failures, alert, and save inputs for later replay. N8N allows to add retries with backoff with many nodes.
N8N limitations
The cons of using N8N is just the other side of the coin of it being a low-code environment:
- You are limited to the nodes and parameters provided. Although you can add custom code nodes, it’s harder to build very complex or unusual logic. For AI functionality N8N nodes are still fairly shallow wrappers — they cover basics like calling OpenAI, embeddings, or a LangChain node itself. They don’t expose advanced features like custom memory classes, structured output parsers, tool-calling orchestration or dynamic retriever pipelines. So you will have to use LangChain or code nodes — which kind of defeats the whole purpose of visual flow editing.
- Workflows run inside its orchestration engine. For large-scale data or very high-performance needs, it may become slower or resource-intensive.
- Workflows are stored in a database/UI format. While you can export JSON, it’s not as natural for versioning and team collaboration so you can’t easily use Git and other tools for smooth collaboration, code reviews, and CI/CD pipelines. n8n offers Git-based Source Control for workflows in paid/self-hosted enterprise and cloud tiers but again it’s not ideal. As you can imagine it can be a no go for a team of any substantial size.
- Debugging can be tricky. You mostly see node inputs/outputs, but you don’t have full insight into the runtime internals, which for any complex project becomes a necessity.
Overall, as you can see it is not ideal for complex projects with many team members involved. It can be used in such scenarios but will likely degrade into using a lot of custom code, avoiding which is the purpose of the framework.
