n8n in 2026: self-hosted vs Cloud, real costs and when to hire help

For a company in 2026, self-hosted n8n versus n8n Cloud is three lines on a page: the license (zero on Community Edition, or a paid plan), where the data sits, and who gets up when the workflow stops. Cloud Starter, on the official pricing page as read on 2 October 2026, is €20 a month billed annually, with 2,500 executions. Free self-hosting charges no license and charges for a server, backups, and a person. Hiring help makes sense when nobody inside owns the credential and the failure.

The figures below are n8n's prices on that date, plus one public server price and, for buyers paying from Brazil, the central-bank euro rate from the same day. They are not a market average. If you want someone to run the workflow, the service is n8n automation for companies.

What the pricing page actually shows

The page has a monthly switch and an annual switch. The 2 October 2026 reading was on "Annually (Save 17%)". Every card price below is the monthly equivalent of the annual plan. The month-to-month sticker, without the 17%, was not on the screen. Treat every "per month" number here as the annual plan spread across twelve months, not as the invoice of someone who pays monthly Source [1].

An execution is one full run of a workflow, start to finish, however many steps it has. n8n uses that against tools that bill per step. To estimate, count triggers. A daily workflow spends about 30 executions a month. A workflow every 5 minutes spends about 8,600 to 8,900. In the chatbot example on n8n's own FAQ, you multiply conversations by messages, because each message is an execution Source [1].

The comparison table on the same page does not match the Pro card on every cell. The card says "up to 50" concurrent executions. The table says "20 / 50". The Assistant-credits FAQ says "4,000 or 9,600 on Pro, depending on your plan size." The card says "up to 9,600." Use the number in the contract you sign, not only the card Source [1].

Other limits the table showed for Starter and Pro: workflow history of 1 day on Starter and 5 days on Pro. Saved executions: 2,500 and 25,000. Storage for those executions: 2.5 GB and 25 GB. Log retention: 7 days and 30 days. Maximum execution length: 5 minutes on Starter and 40 minutes on Pro. Users: unlimited on the plans listed. The Business column was not readable in that rendering. Do not invent a Business limit by copying Pro Source [1].

Business overage: workflows keep running. If you do not upgrade, the overage is invoiced 45 days later, at €4,000 per extra bucket of 300,000 executions. There are no pro-rata refunds. Invoice and wire transfer exist for annual Business and for Enterprise. The Cloud trial does not ask for a card. The Business trial does, and it lasts 14 days Source [1].

Startup plan: a company under 20 employees can get 50% off Business if it qualifies. Community Edition is the standard self-hosted build on GitHub, without that card price Source [1].

Data hosted by n8n sits in the EU, in Frankfurt. Self-hosted data sits wherever you put the server Source [1].

The self-hosted license is not "open source, do what you like"

The LICENSE.md in the repository, Sustainable Use License version 1.0, puts the limit in one sentence: you may use or modify the software only for your own internal business purposes, or for non-commercial or personal use. You may distribute it only free of charge, for non-commercial purposes. A file with .ee. in the name is not under this license. It is under the Enterprise license Source [2].

The license FAQ applies to self-hosted n8n, not to Cloud. "Internal business purpose" includes people outside the business receiving or seeing what the workflows produce. Allowed: run your business, learn, build automations for a client on your instance as long as the client cannot create or edit them, charge for creation, setup, maintenance, training, and consultancy, use n8n behind a product where users only trigger what you built, run several instances Source [3].

Not allowed: host n8n so clients build workflows, let end users configure workflows through a custom UI, the API, MCP, or an AI agent acting for them, fork the code to launch another automation product, remove the branding and resell it, use an Enterprise feature without the license Source [3].

Installing n8n on the client's own server is allowed if you are not also hosting it for them. A partner program for managing many instances is described as in development. The published contact is license@n8n.io Source [3].

Community edition includes almost the whole product. The docs list what it leaves out: custom variables, environments, an external secret store, external storage for binary data, log streaming, multi-main mode, projects, SSO, sharing of workflows and credentials, and Git version control. The same page says queue mode is included in Community, and multi-main is not. The pricing page, which those docs tell you to treat as the current feature breakdown, lists "Queue mode for scaling" on the Business card Source [1] Source [4]. If your decision hangs on queues, read both pages on the day you buy. Do not assume the card and the doc say the same thing.

Registering Community with an email stays free and unlocks folders, debug in editor, and custom execution data Source [4].

What a server actually costs

Community license: €0. The money leaves through the server.

One public, dated price, not a marketing sticker: Hetzner's price adjustment of 15 June 2026, page last changed on 8 July 2026. For new orders and for cloud instance rescales from 8 AM CEST that day, the CX23 cloud server in Germany (FSN/NBG) and Finland (HEL) is €0.0088 an hour, or €5.49 a month, excluding VAT and excluding IPv4. The same row in dollars: US$0.0104 an hour, or US$6.49 a month, also excluding VAT and excluding IPv4. The previous price on that row was €3.99 / US$4.99. IPv4 is billed separately and is not inside the €5.49. VAT is not inside it either. The page points to the German list for prices that include 19% VAT Source [7].

That is an example of a small VPS, not a sizing recommendation. A workflow that moves large files, or that runs a queue and several workers, wants a different machine. Backups, a domain, and the hours of the person who applies the upgrade are not inside the €5.49.

If the card is Brazilian

Buyers in the US and Europe can ignore this section. Buyers paying n8n or a European host on a Brazilian card cannot.

The Central Bank of Brazil's PTAX closing rate for the euro on 2 October 2026, bulletin at 13:03:16, was a selling rate of 5.88150 reais per euro. The buying rate was 5.88030 Source [8]. The arithmetic below uses the selling rate.

The formula is euros × 5.88150. The IOF estimate is 3.5% of that result, rounded to the cent. The card issuer's own FX spread is not in the table. The statement will differ.

The 3.5% figure is the rate written in article 15-B, item VII, of the compiled text of Decree No. 6,306 of 14 December 2007, as amended by Decree No. 12,499 of 11 June 2025. That item covers foreign-exchange operations that settle obligations of institutions in cross-border payment arrangements, as issuers, arising from their users' purchases of goods and services abroad. That is the international-card case Source [9].

The same lines of the compiled text carry two notes: "see Legislative Decree No. 176 of 2025" and "see ADC No. 96." Legislative Decree No. 176, of 26 June 2025, published in the official gazette on 27 June 2025, suspended Decrees No. 12,466, 12,467, and 12,499 and restored the earlier wording of Decree No. 6,306. The legislative decree itself points at ADC No. 96 Source [10].

The table uses 3.5% because that is the rate printed in item VII of the compiled text Planalto publishes. It is not an opinion that the dispute is over. Read the percentage your card issuer actually charged. Do not treat an invoice or a wire as "no IOF." That is a different foreign-exchange operation, and this page does not measure it.

What breaks in production

Credentials and the encryption key

On first launch, n8n generates a key and stores it in ~/.n8n. The key encrypts credentials before they are written to the database. You can set it with N8N_ENCRYPTION_KEY, and the docs tell you to. In queue mode, every worker needs the same key as the main instance. Without the key, a restored database or an encrypted credential export cannot be decrypted Source [6].

A complete backup is the .n8n folder (config, the key, and the SQLite file if that is the default database) plus external Postgres if you use it, plus external binary storage and custom nodes. With SQLite, stop n8n before you copy the folder. Copying while it is writing can produce an inconsistent backup. If you cannot stop it, take a consistent snapshot Source [5].

n8n export:workflow --backup and the credentials equivalent export workflows and credentials as separate JSON files. They do not include users, roles, execution history, variables, or the key. Imported workflows come in deactivated. The flag that restores the exported active state works only in queue mode or multi-main. An export with --decrypted writes passwords in plain text. Delete that file when the restore is done Source [5]. n8n recommends a full backup before you update Source [5].

Versions, queues, and what Community does not include

Upgrading without the key written down is the classic incident. The new container starts, the old volume did not come with it, the credentials are in the database, and they do not decrypt. Write the key down outside the server, in a vault, before the first upgrade.

Queues are the second incident. The editions doc says queue mode is in Community and multi-main is not. The pricing card puts "Queue mode for scaling" on Business Source [1] Source [4]. A worker without the same N8N_ENCRYPTION_KEY cannot read the credential Source [6]. Do not turn a queue on at launch without having run a worker once.

SSO, separate environments, Git version control, projects, and an external secret store are not in Community. They are on the paid plans Source [4]. Buying a build sprint does not flip those switches. The license is a separate invoice, from n8n.

The pricing page lists "Error workflows" under debugging. The table does not make clear which plan includes it. Do not count on a ready-made alert until you see the item on the plan you bought. On self-hosted, logging exists in Community. Log streaming out of the instance does not Source [4].

A Business or Enterprise license on a self-hosted instance must contact n8n's license server every day. The ping includes production execution counts. Telemetry is on by default and can be turned off Source [1].

WhatsApp numbers, including the Brazilian extra digit

If WhatsApp is a side channel, you can skim this. If WhatsApp is how the customer buys, in Brazil, Mexico, Spain, Israel, or much of Latin America, this is the production bug.

Meta's business-phone documentation says to include a plus sign and the country calling code. Omit the plus sign and the API prepends the business number's country code. The message can be misdelivered. The explicit note: for Brazil and Mexico, the extra prefix on the phone number may be modified by the Cloud API. Meta calls that standard behavior, not a bug Source [12].

A workflow that copies 11987654321 or (11) 98765-4321 out of a spreadsheet and sends it will fail, or it will reach the wrong person. Normalize to E.164, with +55 for Brazil, and test with a real mobile number before you arm the trigger. Do not guess which area codes drop the extra 9. Meta's page does not list them. It says the prefix may be changed.

The extra digit itself is not a WhatsApp invention. From 14 February 2017, every Brazilian mobile number gained a 9, in the form 9XXXX-XXXX, by an Anatel rule. Anatel's page was published on 2 March 2015 and updated on 8 December 2020. A caller adds the 9 in front of the old number Source [11]. The CRM often stored the number the way a person dials. The API does not.

Two other blocks on the same Meta page show up in go-live week. A number already on WhatsApp Messenger has to be deleted there before API registration. Registration requires a two-step verification PIN. A new portfolio is capped at two numbers until the business is verified or it reaches a messaging limit of 2,000 Source [12].

An agent that talks on that channel, rather than only firing a template, is covered under WhatsApp AI agents.

Freelancer, studio, or an internal team

A freelancer is the right shape for one well-specified workflow, on systems you can already open, with someone at the company who will own it afterward. The deliverable is the workflow. Monday is yours.

A studio is the right shape when the workflow has to run in production and that owner does not exist: credentials, WhatsApp, a failure someone can see, an upgrade that does not lose the key, more than one system. At samambai a sprint is US$1,200 for two weeks, and you can cancel when you want. Most automations go live in two to four weeks. A fixed-scope project starts at US$3,000 and usually takes four to eight weeks. An ongoing partnership starts at US$900 a month. The first conversation is 30 minutes and does not commit you. The three shapes are on pricing.

The studio is the wrong shape in four cases. The workflow is a single step and you already operate n8n. You want your customer to build automations, which the community license does not allow Source [3]. You want the data to live in the builder's account. Automation is the company's product and the volume already calls for an internal on-call engineer plus a Business or Enterprise license, which you buy from n8n, not from the sprint Source [1].

An internal team is the destination when the workflow is the business: a lot of executions, SSO, a test environment that is not production, someone on call. Before that, an internal team with no on-call owner only moves the place where the key gets lost.

The studio runs its own n8n server, with dozens of production workflows across support, integrations, and routines. That is not a count of messages saved. It is the kind of operation: the key exists, the failure shows up, the upgrade has a backup first. The studio is based in São Paulo and already works with clients in São Paulo, New York, Madrid, and Tel Aviv, in English and in Portuguese.

Client data stays in the client's accounts. Access is the minimum required. It is not used to train public models. While a workflow is built, AI agents help write and test it, and a person reviews it before it is activated.

If the decision is still Cloud or a server, do the arithmetic with your own trigger count and with the official table, not with last year's price. If the decision is already "it has to run on Monday," the next page is n8n automation.