n8n

You need to consider two things: whether you have a server or someone who can set one up, and whether you are ready to work in an English interface.

What people come to n8n for and what they get: six common questions where the answer is not obvious.

Who it is for. For those who already have a repetitive task: put website leads into a CRM, send a summary to Telegram every morning, reply to customer emails with a neural network. And for those who want to build an AI agent without glue code: an AI Agent node, plus a model, memory, and tools, all connected on the canvas in an evening.

Who else. Anyone who keeps data on their own machines. A self-hosted install is free, does not count executions and never calls a license server.

Who it is not for. Anyone waiting for a "press a button and you're done" tool. The interface is English, the first hour goes on Docker and a domain for webhooks, and for ease of use n8n scores 4.0 out of 5 on Capterra against an overall 4.6.

Good fit: section 3 takes you from a server to your first workflow. If you're choosing between n8n and Make, see the choice in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

Three layers: ready-made nodes, your own code inside a node, and an AI agent that decides for itself which nodes to call.

Nodes. As of Sep 6, 2026 the catalog shows 2,147 nodes: Telegram, Google Sheets, Gmail, Slack, Airtable and an HTTP request to any API. The number of integrations differs from one vendor page to another, from "400+" in the repository description to "1500+" in the README, so counting by the catalog is more reliable.

Code. You can write JavaScript or Python in the Code node. npm packages can be added to a self-hosted installation; external libraries are unavailable in the cloud, and Python in the cloud cannot import anything at all. The catalog has 12,125 ready-made workflow templates written by the community.

AI agent. The AI Agent node has four connectors: model, memory, tools, and response parsing. You can connect any model, from OpenAI and Anthropic to DeepSeek and local models through Ollama, while the Model Selector node can distribute requests across four models from different vendors. The agent can access external MCP servers, and the workflow itself can be exposed as an MCP server; since Aug 10, 2026, you can connect to 70 MCP servers from the node panel via OAuth.

Build with words. AI Workflow Builder creates a workflow from a description such as «every morning at eight, collect AI news, summarize the five biggest stories, and send them to Telegram», and a chain of a dozen nodes appears on the canvas. Every edit through the builder costs one credit from the cloud plan's monthly allowance; a self-hosted installation needs your own model key.

What the vendor itself shows. The video in the header comes from the n8n.io homepage: on a dark canvas, Schedule Trigger, Find Comments, HTML Extract, Edit Fields, and Code nodes appear one after another and connect with green links. We did not run n8n for a specific task ourselves; the figures come from documentation, pricing pages, the GitHub API, and third-party tests.

Install it on your own server through Docker: the npm install stops being supported in version 3.0 in October 2026.

Step 1. Take a server with Docker and at least 1 GB of memory: n8n itself eats 190 MB, and the vendor's entry cloud plan lives on 335 MB.

Step 2. One command from the vendor website will start the container itself:

curl -fsSL https://get.n8n.io | sh

Or use the image directly; the editor will open on port 5678:

docker run -it --rm --name n8n -p 5678:5678 -v n8n_data:/home/node/.n8n docker.n8n.io/n8nio/n8n

Open localhost:5678 or the server address, then create the owner account by email. A free license key sent to the same email adds folders and debugging to the editor.

Step 3. For a Telegram bot and any webhooks, n8n must be accessible from the internet: you need a domain and an HTTPS certificate. For local testing, a cloudflared tunnel is enough; the vendor provides it for testing only and explicitly does not recommend it for production.

Practitioners take DeepSeek or OpenRouter, both with built-in nodes, or a local model through Ollama, and then no keys and no payment are needed at all.

Cloud. Cloud data sits in the EU on Microsoft Azure.

Steps verified on Sep 6, 2026 using the repository README, n8n documentation, Timeweb Cloud, Beget, and Amvera pricing pages, and a check-host test with control nodes

Automation builders have no arenas or benchmarks, so the measurements here are about scale and demand, not quality.

Code and community. 203,515 stars and 60,579 forks on GitHub (API, Sep 6, 2026); the vendor places the repository among GitHub's top fifty. The forum has 175,080 registered users and 6,704 active users per month according to the forum API, while the homepage rounds the community to «200 thousand». .

Company funding. On May 12, 2026, SAP took an equity stake at a $5.2 billion valuation, twice the October valuation; the company has raised $240 million in total. In the same press release, n8n reports 1,400 enterprise customers and 1.7 million active builders per month; a year earlier it counted «3,000 enterprises», so the metric changed, and the company does not disclose revenue.

Demand. The place among automation tools is shown under the card title, and the full tool ranking lives on a shared page.

A self-hosted installation costs 0 € and does not count executions, while the cloud sells executions in bundles: Starter is 24 € per month for 2,500, although the pricing page shows the annual price of 20 € by default.

Plan currency. A run in n8n counts for the entire workflow, “no matter how many steps and data it contains,” as the vendor explains on the pricing page. In Make and Zapier, the counter ticks for every operation, so one request to a Telegram bot with an AI model inside costs one run in n8n, but five or six operations with its competitors. The vendor itself calculates on the pricing page: a workflow on a five-minute schedule consumes 8,600–8 900 runs per month, chat-bot runs should be counted by number of messages, conversations multiplied by the average number of messages per conversation, and according to the documentation, manual editor runs and nested sub-workflows do not count toward the quota.

Self-hosting. The Community edition is free, needs no key and counts neither users nor executions. It has no single corporate sign-on (SSO), no projects, environments, Git, external secrets or shared access to workflows: someone else's workflow is visible only to its author and the owner. A free key sent by email adds folders, debugging and execution tags. The Business plan at €800 a month is sold for self-hosted installs too, and its key calls the vendor's license server once a day.

Cloud. Free provides 50 runs per month, Starter 24 € for 2,500, Pro 60 € for 10,000, Business 800 € for 40,000, and Enterprise on request. Annual billing is 17 % cheaper, and those are the prices shown first on the pricing page. The card-free trial is limited to 1,000 runs; the vendor does not state the number of days. The vendor offers no refunds for annual subscriptions, while Business overages are sold in bundles of 300,000 runs for 4,000 €.

Quiet price increase. The figure of 20 on the pricing page has not changed since January 2023; what it includes has. In 2023, 20 $ bought 5,000 runs; from March 2024, it bought half as many; in September 2025, dollars changed to euros and the Business plan appeared. The free plan was introduced in the first half of 2026.

Three questions: what the license allows, how leaky 2026 was, and where n8n sends your data.

License. The vendor calls the model fair-code and explains it with examples. You can synchronize your company's data, write your own nodes, charge clients for workflow setup, and run n8n on an internal server. You cannot sell access to your n8n for money or release it under your own brand. Files marked .ee in the repository are not covered by this license at all; they can only be used with an Enterprise key.

Vulnerabilities. Five critical vulnerabilities were disclosed from Dec 19, 2025 to Feb 4, 2026, and two received names. Ni8mare, scored 10.0 out of 10, provided access to server files without login and had already been fixed in version 1.121.0 on Nov 18, 2025, 50 days before disclosure. N8scape allowed an escape from the Python sandbox and was fixed by version 2.0 itself, which the vendor released on Dec 8, 2025 as “secure by default”. Cloud was affected too; this was not only a self-hosting problem.

Public instances. On Aug 5, 2026, GitGuardian found 4,576 leaked n8n keys in public commits; 321 live instances accepted someone else's key, while internet search revealed 31,793 installations. Since January 2026, n8n has had 48 vulnerability records by the same company's count. An instance stores keys for dozens of services, so updates and restricted access matter more than any other setting.

Data. Cloud runs on Microsoft Azure in the EU, and the vendor undergoes an external SOC 2 security audit once a year. Self-hosting sends data nowhere except for one thing: by default, it sends usage telemetry to the vendor, which can be disabled according to the instructions. A paid key pings the license server every day, the keyless Community edition does not contact the vendor's server, and free registration also issues a key. The workflow builder sends the model a description of nodes and connections, but not your keys or previous runs.

Reviews. Product Hunt: 4.81 out of 5 from 73 reviews and a “best for developers” award for winter 2026 (Sep 6, 2026); Capterra: 4.6 from 48 reviews, with 4.0 for ease of use (page data as of Aug 20, 2026). The most common downside in Product Hunt reviews, four mentions out of 73, is a steep learning curve. Practitioners on Habr add their own pitfalls: an 18 GB run database in four months and an out-of-memory crash for one client (May 11, 2026).

The split is not in the number of integrations, but in two questions: where they meter the money, per step or per run, and where the workflow lives, on their side or yours.

If a request goes through several departments and needs a shared approval order, compare the scenario with the Pega platform: it designs the application and runs the cases separately.

Make. A workflow builder from the Celonis group, which grew out of the Czech Integromat: 3,000 ready-made apps, workflows are drawn with modules on the same canvas, and since 2025 can also be built by chatting with Maia. It only runs in the cloud, gives 1,000 free credits a month with runs no more often than once every 15 minutes, Core costs 12 $, Pro 21 $ for 10,000 credits (make.com, Sep 6, 2026). A credit is charged for every module action; custom JavaScript or Python code costs two credits per second.

Zapier. The largest catalog, 9,000 apps, and the simplest model: an event in one, an action in another. Free includes 100 tasks per month; Professional starts at 19.99 $ for 750 tasks with annual billing and 29.99 $ month to month, with every workflow step and every AI call consuming a task (zapier.com, Sep 6, 2026). No self-hosted server, less code inside workflows than n8n, but its own MCP server, through which AI agents access the same app catalog.

In seven years, a Berlin special-effects specialist's night project grew to a 5.2 billion $ valuation and gained SAP as an investor, while along the way receiving its own license, two major versions, and a wave of critical vulnerabilities.