@n8n_io killed its lead gen target and won't do per-seat pricing. @JanOberhauser's explanation turned into something better, which is what burning money quietly does to your roadmap. Most AI companies are growing fast and losing a lot. That means you need growth to raise the next round, which pushes you into thinking in quarters. More ARR now so there's more money later, then do it again. Eventually your metrics are about the fundraise rather than the product. n8n is profitable. The next round is optional, so the next quarter isn't a deadline. What he bought with that is permission to chase usage instead of revenue. His question is what actually creates value for customers and how much of the opportunity they can reach, not how fast the number climbs. Per-seat pricing is the clearest example. It's the default for good reasons, and it also punishes the exact thing you want, which is more people inside a company using your product. Charge per seat and your best customers pay for succeeding with you. Most companies take that trade anyway, because investors understand the model. Being able to say no to it comes from not needing them.
n8n on X
The largest part of the talk about n8n is people showing it working: 86 report live builds and results, ahead of 31 praising it in passing. Complaints are narrower: green runs hiding wrong work, and the upkeep workflows take.
First written 27 Oct · what changed · from 293 posts by 230 people
Tracking posts that say “@n8n_io” or “n8n”
Track this for your marketPosts per day, last 30 days
as of Nov 3, 11:01 UTCWhat changed
as of 27 Oct · checked once a monthFirst written 27 Oct. When what people say changes, it shows here.
Tone· 169 posts from 139 people· written 27 Oct
How do people feel about n8n?
Most of the conversation is people reporting n8n working: 86 people describe builds they run for real work, and 31 more praise it outright. Scepticism is the minority — 8 people say the buzz has faded and 5 say they tried it and moved off.
What people describe is concrete: a sneaker-cleaning business run from Telegram with Supabase and a few SQL functions, leads scored out of 100 with a hot-lead email inside five minutes, an invoice parser that sends every bad download or AI error to its own log. Several count scale — 500+ workflows, five agents running 24/7 — and many say self-hosting is where the value is: the free Community Edition, a NAS at home, and one who finds a $20-a-year price makes trying it almost frictionless.
A smaller group praises the shape of the product rather than one build: agents with their own durable state, 70+ native AI nodes, memory that carries across runs. One poster calls it "underrated for production automation".
11 people single out keeping a person in the loop: approval before anything sends, a booking flow that re-checks the calendar and then drops messy replies into Slack for a human.
Against that, 8 people say the buzz is over and the tool feels "a thing of the past", while 5 describe leaving for code after workflows proved hard to edit or a webhook cost them a night.
- Satisfied in practice109
- Praises it35
- Human in the loop is the right way11
- Says the hype has faded8
- Tried it and moved off6
See 5 of 169 postsHide the posts
Catching up on what @n8n_io is doing with agents and it’s definitely interesting. Agents are now first class and have their own durable and resumable state. Workflow steps can send messages to those agents and give them tasks. Agents can invoke workflows as tools (not completely new but improved). link
My sneaker cleaning business now runs from Telegram. "list" → every open order "ready 12" → order 12 marked ready "done 12" → closed And every morning at 7 AM it sends me what's overdue, before I leave for my 9–5. No dashboard to open. No spreadsheet. Just a chat. n8n + Supabase + a few SQL functions.
Just finished the GenAI Architect Program by @SocialEagle (Aug 24 → Oct 2, 2026). 🎓 6 weeks ago my work was WooCommerce stores. Now I build AI agents. Here's what I actually learned, beyond the certificate 👇 1. RAG is mostly a retrieval problem, not a model problem. When a RAG system gives bad answers, the LLM is rarely the cause. It's chunking, metadata, and whether the right context reaches the model at all. Fix retrieval first. 2. n8n is underrated for production automation. Not every workflow needs custom code. Connecting a CRM, inbox and store with an LLM step in the middle solves a surprising number of real business problems. 3. Multi-agent ≠ better by default. Splitting work across specialised agents is powerful, but each handoff adds a point where things can fail. Start with one agent. Split only when it clearly struggles. 4. Voice agents are where the "wow" is. Building with LiveKit showed me how close we are to agents handling real phone conversations. Latency is the whole game. 5. Fine-tuning is the last resort, not the first. Prompting + RAG covers most use cases. Fine-tune when you need consistent style or behaviour at scale. 6. Claude Code changed how fast I ship. Pair it with a solid CLAUDE.md per project and it stays consistent across clients. The unexpected advantage: my 5+ years in QA. AI systems don't crash. They fail quietly, with confident wrong answers. Testing edge cases, writing evals and asking "how could this break?" matters more in AI than in almost any software I've worked on. What's next: 🚀 Building AI agents & automations for businesses at Digital Pro Creators 🛠️ Building my own MicroSaaS with Claude Code, in public. Idea selection, build, launch, first paying user. Real numbers, real mistakes, all shared here. If you're building in AI or MicroSaaS, let's connect. Follow along for the journey. 🤝 #BuildInPublic #AIAgents
Salon WhatsApp fills with Thursday at 3 while two clients are already fighting for the same chair. My n8n booking flow re-checks Google Calendar on the confirm step, then dumps the messy replies into Slack for a human.
5 of the 169 posts behind this answer, one per person. The other 164, and every new one as it’s posted, come with tracking. Get early access
Pain points· 66 posts from 53 people· written 27 Oct
What do people complain about with n8n?
The complaint people make most is that a run can look fine and still be wrong: 13 people describe green runs hiding skipped items, duplicates and work that never happened. Next, 11 describe how much effort building and keeping workflows takes, and 10 raise security worries. 8 complain about cost and billing.
A green run only means nothing threw: the Claude API returns a polite refusal instead of a post, a filter step drops every item after a field is renamed, a step passes because it only checked that SSH connected. One workflow waited eleven times for a video that would never exist, because the provider's success reply carried a 402 inside it; another saw Stripe resend an event into a duplicate model call.
Effort comes next: the interface is called "codish" and hard to navigate with many nodes, and one asking AI for a workflow got a trigger plus 400 lines of JS in a Code node. Others describe three hours on one "straightforward" automation and a webhook that cost a night to find.
Security talk is about exposure: encryption keys in env vars, credentials in exported JSON and pasted into HTTP headers, custom nodes, 76% of versioned hosts in a CVSS 9+ range.
Money: a trial used up in a weekend, credit running out mid-workflow, per-run billing on frequent schedules, 48 hours waiting for a billing answer.
Smaller groups fight self-hosting setup and licence limits, report updates breaking nodes or slowing batch jobs, and say durable agents still lack approval gates and audit trails.
- Runs fail silently or wrongly18
- Hard to build and keep working14
- Security worries10
- Costs, credits and billing8
- Durable agents need more control6
- Self-hosting and deployment pain5
See 5 of 66 postsHide the posts
In link and n8n, a green run only means nothing threw an error. That is a much weaker promise than it looks. An AI content pipeline fails quietly far more often than it fails loudly. The Claude API returns a polite refusal instead of a post. A filter step drops every item because a field got renamed upstream. The scheduler publishes an empty string because a variable mapped to nothing. Every one of those runs ends green. So silent failure is not an edge case. It is the default state of any automation that only checks whether its steps executed. The fix is to stop asking 'did it run' and start asking 'did it produce what it should.' That means assertions on output, not on execution. Text below a minimum length fails. Text containing typical refusal phrasing fails. A run that yields zero items when it normally yields several fails. When an assertion trips, the run should go red on purpose. A loud failure gets fixed. A quiet one ships. The tradeoff is real. Strict checks throw false alarms, and a pipeline that cries wolf gets muted. Loose checks let garbage through. What works for me: fail hard on empty or malformed output, only warn on anything merely unusual, and review the warnings weekly to tighten or loosen the rules. What is the quietest failure your automation ever shipped before anyone noticed?
2/ An n8n step ran a command over SSH that didn't exist on the remote box. The remote returned an error. The step passed anyway, because it only checked that SSH connected. Now any non-zero exit fails the run.
An AI video provider ran out of credits, and my n8n workflow kept waiting for a video that was never going to exist. Its wait step ran 11 times, and it would have kept going forever.
At first I found @n8n_io really "codish" I found the interface quite difficult to navigate. All those nodes, connections, and fields? Was a lot And now? I'm getting the hang of it, and it's becoming interesting🥹 What's one thing you are learning?
Me: This automation is very straightforward. I should be done in 30 minutes. Also me, three hours later: Why is this field an array? Why is this expression returning undefined? Who changed the data structure? Why is the IF node behaving like this? 😭 And the funniest part? After finally fixing everything, I will still open n8n tomorrow and start another workflow. I think I enjoy suffering in JSON format. 😂
5 of the 66 posts behind this answer, one per person. The other 61, and every new one as it’s posted, come with tracking. Get early access
Alternatives· 36 posts from 32 people· written 27 Oct
What are the alternatives to n8n?
The main thing people weigh n8n against is writing the code themselves: 12 people argue for their own scripts, backends or Claude Code over node workflows. Zapier, Make and Claude are level behind at 4 people each.
The code camp splits in two. Some left after a bad experience: workflows hard to edit, schedules overcomplicated with schemas, a webhook that broke for a night, one now running the same customer-service system as a Python file and a markdown doc. Others argue the shape is wrong, calling visual node workflows archaic or n8n "just stupidly hard to use". One gives it its due — n8n is not dead, the use case shifted, and a workflow tool still wins for recurring operations with retries, credentials and handoff to a team.
Price is the usual ground against Zapier: per-task billing compounding on agent workflows with hundreds of micro-calls, "5-8x cheaper at scale", and a self-hosted copy presented as $0 with unlimited executions against a ten-step flow burning ten tasks. The trade offered is fewer integrations and a server you own.
Make is weighed on value for local businesses and complex visual logic; Claude on scheduled tasks that replace whole workflows, and on building an n8n workflow in five minutes. Smaller rivalries — Grok Bot, Power Automate, OpenAI Agent Builder, PhantomBuster, Clay, SAP, Langdock, Mastra — each rest on one or two people.
- Coding it yourself13
- Zapier6
- Make4
- Claude instead of n8n4
- Grok Bot2
- Power Automate1
See 5 of 36 postsHide the posts
n8n got replaced by code.. and honestly they complicated too much with the schemas
Honestly been migrating a lot of stuff away from it lately. Found it much better to use AI to build my own backend, most of my n8n projects had custom code in it. Just made more sense when managing latency to move away and build on my own stack. The managed service sucked a
ill say it then, ill say it now, i would rather ship code and build systems using claude code (md files, cron jobs, yata yata) but ill never use n8n. it's just stupidly hard to use
n8n is not dead, the use case just shifted. One-off scripts are fine for prototypes and small projects. For recurring ops with retries, credentials and handoff to a team, a workflow tool still wins. Python replaces glue code, not bigger operations.
The cost gap is not subtle. Zapier bills per task. In an AI agent workflow with hundreds of micro-calls, costs compound fast. n8n runs 5-8x cheaper at scale — and self-hosted is effectively free. That alone changes the economics of building with AI.
5 of the 36 posts behind this answer, one per person. The other 31, and every new one as it’s posted, come with tracking. Get early access
Questions· 34 posts from 31 people· written 27 Oct
What do people ask about n8n?
The largest group of questions is operational: 9 people want to know how others run and debug their own setups — kill switches, which prompt handles a branch, which setting cost the most time. Behind them, 8 ask whether n8n is still used.
The debugging questions are specific and practical. Someone wants to know what a client can flip to stop an agent without calling them, and settles for a toggle in a Google Sheet. Others ask which of three prompts handles the branching logic on a dense canvas, and what the one setting is that cost the most time while debugging. One asks what the dumbest non-AI thing is that broke an agent, offering a cron running in UTC that fired reminders four hours early and an API key that expired on a Friday night.
The second question is whether the tool still matters. People ask "Does anyone still use n8n?", notice the $9999 workflows have dried up, or ask what happened to it. The answers posted back say it is fine and being used by companies, with the hype the thing that died.
Smaller questions follow: whether to learn n8n or just code, how to make a flow safe with failure paths and approval steps, what the real failure rate is behind published run numbers, and one-off checks on self-hosting, a TikTok node and whether an API key works with it.
- How do you run and debug yours?10
- Is n8n still used?8
- How do I make a flow safe?4
- Learn n8n or learn to code?3
- What's the failure rate at scale?3
- What is n8n?2
See 5 of 34 postsHide the posts
every agent you hand a client needs an off switch they can flip without calling you a toggle in a google sheet counts. honestly better than most dashboards what does your kill switch look like? @n8n_io
That n8n canvas looks dense — curious which of the 3 prompts handles the branching logic visible in the workflow.
every time i'm sure the model broke a client agent it ends up being plumbing last one was a cron running in UTC so reminders went out 4 hours early. before that an api key that expired on a friday night what's the dumbest non-AI thing that broke your agent? @n8n_io
Does anyone still use n8n?
I haven’t seen any posts about "$9999 value n8n workflows" for a long time. What happened to the n8n crowd? Are they still around? I was thinking about learning n8n, but I never actually got started. Just as well, I suppose.
5 of the 34 posts behind this answer, one per person. The other 29, and every new one as it’s posted, come with tracking. Get early access
Wishes· 7 posts· written 27 Oct
What do people want from n8n?
There is no shared wish here: 7 people each ask for something different, so what people want from n8n is told one request at a time rather than as a pattern.
The asks around agent building are for control. A native ZoomInfo node, so search and enrich are tools an agent can call without a separate integration hop. Routing logic kept apart from text generation, so failures are caught deterministically before touching downstream nodes. A chat that explains why a workflow routed one item one way and not another, raised as part of a longer account of the real cost of entry.
For the voice and photo workflows, the asks are about what happens when input falls short: a spoken "couldn't find that place" when a city lookup fails, and a short input guide for the five-photo workflow — varied angles, consistent identity, no conflicting accessories — plus a fixed approved reference set when regenerating just one failed shot.
The rest are small product fixes: table views people can save, because they like working in n8n's tables, and a square version of the new logo, because it looks bad in Discord thumbnails.
- ZoomInfo integration1
- Save table views1
- Chat that explains decisions1
- Square logo1
- Photo workflow input guide1
- Spoken fallback for missing data1
See 5 of 7 postsHide the posts
Native ZoomInfo inside n8n is the piece I'd test first for agent workflows: search and enrich as tools the agent can call without a separate integration hop.
Running the decision engine and generation models locally solves the compliance barrier for enterprise pipelines.The real leverage comes from separating the routing logic from text generation so failures can be caught deterministically before touching downstream nodes.
Problem jest to ze trzeba wydać pieniądze ;) Jeżeli to faktycznie ma działać lokalnie, to jaki najtańszy sprzęt się nada, jaki jest realny koszt wejścia i po jakim czasie będzie faktyczny zwrot z inwestycji, pozostaje jeszcze kwestia zbudowania zaufania do takich rozwiązań, nawet jak działają lokalnie :) Fajna opcją do takiego workflow byłaby jeszcze opcja czatu który mógłby wytłumaczyć dlaczego np dany e-mail trafił tu a nie gdzie indziej, pytania o proces decyzje
What does the caller hear for a city the API can't find? A spoken 'couldn't find that place' would be handy for that case.
A small input guide would help users get more from those five photos: varied angles, consistent identity, and no conflicting accessories. It would also be useful to keep the approved reference set fixed when regenerating just one failed shot.
5 of the 7 posts behind this answer, one per person. The other 2, and every new one as it’s posted, come with tracking. Get early access
In their words
How do people talk about n8n?
The phrases people use most, counted once per person.
- 1“human approval”5
- 2“approval step”5
- 3“durable state”3
- 4“api keys”3
- 5“workflows as tools”2
- 6“custom code”2
- 7“visual workflows”2
- 8“repetitive work”2
- 9“cheap check”2
- 10“rite of passage”2
Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.
Also in Workflow automation
- Zapier+240%Talk about Zapier splits between people describing Zaps that quietly run their leads, inboxes and CRM syncs, and 13 people who think a prompt or an agent now does the same job.206 posts / 7d
- Gumloop−5%Gumloop's talk is largely praise for its Agent Browser — agents that log in and click where there is no API — with a thin, split set of complaints about broad upfront permissions, fragile browser recordings and the $37/mo price.44 posts / 7d
- Lindy−100%Talk about Lindy is warm but general: the biggest group, 20 people, offers plain goodwill, while 15 credit what it digs up across Slack, email and Drive before a launch.11 posts / 7d
Get every post about n8n, filtered and labelled
This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.