n8n, Huginn and Node-RED give you triggers, persistence and unattended execution without a model: a schedule or an event starts a flow, each step is a visible node, and the result is written somewhere. They have done this for over a decade. Add a model at exactly one node, the one that needs judgement, and you have the cheapest always-on agent you will ever run.
Part of the Always-on agents series, post 29 of 33.
In the agent, chatbot or automation post I gave myself one rule: if I can write the decision as a rule, I write the rule and I do not call it an agent. The tools in this post are where those rules live. They are not new, they are not clever, and two of them have been running on other people's servers since 2013, which is the whole point. An always-on job that does not need judgement should run on something that has already proven it can stay up.
Three tools, three licences
| Tool | What it is | Licence | Age and size, read on 20 October 2026 |
|---|---|---|---|
| n8n | Fair-code workflow automation platform with native AI capabilities, in its own words; a visual editor over hundreds of integrations, self-hosted or cloud | Fair-code: distributed under the Sustainable Use License and an enterprise licence, so not open source in the OSI sense | 208k GitHub stars |
| Huginn | Create agents that monitor and act on your behalf, per its README; a self-hosted Rails app the project calls a hackable alternative to IFTTT or Zapier | MIT | Created by @cantino in 2013; 50.1k stars |
| Node-RED | Low-code programming for event-driven applications; a flow editor and runtime on Node.js | Apache-2.0 | Started in early 2013 as a side project at IBM's Emerging Technology Services by Nick O'Leary and Dave Conway-Jones, open-sourced that September, a founding project of the JS Foundation in 2016, now under the OpenJS Foundation; 23.8k stars |
Two of the three are open source in the strict sense. n8n is not, and I think the honest thing is to say so plainly rather than file it under open source because it is on GitHub. For a company running it internally the Sustainable Use License is rarely a problem; for a product built on top of it, read the licence first.
The flow I built in both: watch a feed, keep the matches, flag what needs a human
This is the listening half of the brand watch agent, with no model in the loop except at one node, and it is the same flow in Node-RED and in n8n.
- Trigger on a clock. Once an hour, and once at start so you see it work. Node-RED's inject node has a repeat interval and a fire-once-on-start switch; n8n's Schedule Trigger has an hourly rule.
- Fetch a feed. An HTTP request in Node-RED, parsed with the xml node; n8n has an RSS Read node that returns one item per entry. I used the Hacker News search feed for the word agent.
- Filter by rule. Title contains agent. A function node in Node-RED, a Filter node in n8n. This is a rule, so it is a rule, not a prompt.
- One judgement node, stubbed. Which hits need a human now? In my test it is a regular expression for security, outage, breach and incident. In production this is the one node that calls a model, with the title and link and a one-word answer: human or log.
- Write the result where you will see it. A file in this test; a Telegram message or a webhook in yours. Overwrite, not append, so the file is always the latest run and a stale run is visible.
Node-RED
The flow is a JSON file. Drop it into the data folder, start the container, and the runtime loads it. No account, no login, no editor needed, although the editor is there on port 1880 if you want to see the wires.
nodered/flows.json (trimmed to the fields that matter)[
{"id":"t1","type":"tab","label":"feed watch"},
{"id":"i1","type":"inject","z":"t1","name":"every hour (and once at start)",
"repeat":"3600","once":true,"onceDelay":2,"payloadType":"date","wires":[["h1"]]},
{"id":"h1","type":"http request","z":"t1","name":"fetch the feed","method":"GET","ret":"txt",
"url":"https://hnrss.org/newest?q=agent","wires":[["x1"]]},
{"id":"x1","type":"xml","z":"t1","name":"parse","property":"payload","wires":[["f1"]]},
{"id":"f1","type":"function","z":"t1","name":"filter: title mentions agent",
"func":"const items = (msg.payload.rss && msg.payload.rss.channel[0].item) || [];\nconst hits = items.filter(i => /agent/i.test(i.title[0]));\nmsg.items = hits.map(i => ({ title: i.title[0], link: i.link[0] }));\nmsg.count = hits.length;\nreturn msg;",
"wires":[["j1"]]},
{"id":"j1","type":"function","z":"t1","name":"judgement (stub for the model call)",
"func":"// The one place a model belongs: deciding which hits need a human now.\n// Stub rule in this test; a real flow calls a local model here with the title and link.\nconst urgent = /security|outage|breach|incident/i;\nconst lines = msg.items.map(i => (urgent.test(i.title) ? 'HUMAN ' : 'log ') + i.title + ' :: ' + i.link);\nmsg.payload = lines.join('\\n') + '\\n';\nreturn msg;",
"wires":[["w1","d1"]]},
{"id":"w1","type":"file","z":"t1","name":"write matches.txt","filename":"/data/matches.txt",
"overwriteFile":"true","wires":[[]]},
{"id":"d1","type":"debug","z":"t1","name":"count","complete":"count","console":true,"wires":[]}
]container log## node-red log (filtered)
20 Oct 22:37:37 - [info] Node-RED version: v5.0.8
20 Oct 22:37:38 - [info] Started flows
20 Oct 22:37:41 - [info] [debug:count] 20
## /data/matches.txt
-rw-r--r-- 1 node-red node-red 2493 Oct 20 22:37 /data/matches.txt
log How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering? :: https://arxiv.org/abs/2609.40303
log Show HN: Agent-Zero - Starter template for AI-native coding projects :: https://github.com/amorey/agent-zero
log Ask HN: Are agent benchmarks the new credit ratings? :: https://news.ycombinator.com/item?id=50172009
log Mouse against Opus 5.5 in an agent harness. 13x faster :: https://twitter.com/heypeterjames/status/2112621952879771744
log Show HN: Harness Studio: an open-source desktop coding agent for any model :: https://github.com/vigneshv1cky/harness-studio
log Show HN: What the Harness? LLM agent inconsistencies in MCP implementation :: https://what-the-harness.polytomic.com
log Minlit - an AI agent that reads CS/ML papers and answers with citations :: https://www.minlit.app/ask
log Show HN: Gauge CLI - Open-Source Hillclimbing for Agent Experience :: https://github.com/gauge-sh/gauge-cli
(3 feed titles contained a dash character; shown as a hyphen here, nothing else changed)n8n
n8n also stores a workflow as JSON, and its command line can import one and run it once without touching the editor, which is how I tested it. The node types and versions in the file matter: they are the contract between you and the runtime, and an import that succeeds with a wrong version can still fail at execution.
n8n/wf.json (trimmed){
"id": "aofeedwatch", "name": "ao feed watch", "active": false,
"nodes": [
{"name":"Every hour","type":"n8n-nodes-base.scheduleTrigger","typeVersion":1.2,
"parameters":{"rule":{"interval":[{"field":"hours","hoursInterval":1}]}}},
{"name":"Run once by hand","type":"n8n-nodes-base.manualTrigger","typeVersion":1,"parameters":{}},
{"name":"Read feed","type":"n8n-nodes-base.rssFeedRead","typeVersion":1.1,
"parameters":{"url":"https://hnrss.org/newest?q=agent"}},
{"name":"Mentions agent","type":"n8n-nodes-base.filter","typeVersion":2.2,
"parameters":{"conditions":{"conditions":[{"leftValue":"={{ $json.title }}","rightValue":"agent",
"operator":{"type":"string","operation":"contains"}}],"combinator":"and"}}},
{"name":"Judgement (stub for the model call)","type":"n8n-nodes-base.code","typeVersion":2,
"parameters":{"jsCode":"const urgent = /security|outage|breach|incident/i;\nconst lines = $input.all().map(i => (urgent.test(i.json.title) ? 'HUMAN ' : 'log ') + i.json.title + ' :: ' + i.json.link);\nreturn [{ json: { count: lines.length, text: lines.join('\\n') + '\\n' } }];"}},
{"name":"To file","type":"n8n-nodes-base.convertToFile","typeVersion":1.1,
"parameters":{"operation":"toText","sourceProperty":"text"}},
{"name":"Write file","type":"n8n-nodes-base.readWriteFile","typeVersion":1,
"parameters":{"operation":"write","fileName":"/home/node/out/matches.txt"}}
],
"connections": {"Every hour":{"main":[[{"node":"Read feed"}]]}, "Run once by hand":{"main":[[{"node":"Read feed"}]]},
"Read feed":{"main":[[{"node":"Mentions agent"}]]},
"Mentions agent":{"main":[[{"node":"Judgement (stub for the model call)"}]]},
"Judgement (stub for the model call)":{"main":[[{"node":"To file"}]]}, "To file":{"main":[[{"node":"Write file"}]]}}
}Terminal# Node-RED: the flow file is the whole configuration
docker run -d --name feedwatch --memory=1g -v "$PWD/nodered:/data" nodered/node-red
# flows.json in that folder is loaded at start; the inject node fires once, then hourly
# n8n: import the workflow, then run it once by hand to see the output
# n8n only writes files under a folder you name in N8N_RESTRICT_FILE_ACCESS_TO
docker run -d --name n8n --memory=1g -v n8n_data:/home/node/.n8n -e N8N_RESTRICT_FILE_ACCESS_TO=/home/node/out docker.n8n.io/n8nio/n8n
docker cp n8n/wf.json n8n:/tmp/wf.json
docker exec n8n n8n import:workflow --input=/tmp/wf.json
docker exec n8n n8n execute --id=aofeedwatch # runs from the manual trigger; the hourly trigger needs the server runningcontainer log## n8n version
2.44.5
## import
[license SDK] Skipping renewal on init: license cert is not initialized
Importing 1 workflows...
Successfully imported 1 workflow.
## execute once from the CLI
(n8n prints the whole execution record as JSON here; trimmed)
## output file
-rw-r--r-- 1 node node 2358 Oct 21 14:38 /home/node/out/matches.txt
log Release of Summonr - The open-source coding agent that stays light on your token :: https://summonr.dev/
log Show HN: VT Code - A terminal-first coding agent with sandboxed tools :: https://github.com/vinhnx/VTCode
log Free agent skills that walk you through designing a website with AI :: https://github.com/caelo-cms/website-design-skills
log Show HN: My personal agent, distributed as a prompt :: https://github.com/mattabate/life-distributable
log Show HN: Memdoor - A coding agent that can build and run headless workflows :: https://github.com/guregodevo/memdoor
log Show HN: Mnestiq - tamper-evident logs of what an AI agent saw and did :: https://github.com/mnestiq/mnestiq
log Agent-Head Architecture :: https://agentitter.com/agent-head
log Compression should be part of your agent harness :: https://hackteam.io/blog/why-compression-should-be-part-of-your-agent-harness/
(4 feed titles contained a dash character; shown as a hyphen here, nothing else changed)Huginn, which I did not run
Huginn is the oldest of the three and the one whose vocabulary is closest to this series: it calls its building blocks agents, and has since 2013, long before the word meant a language model. A Website Agent watches a page, an RSS Agent reads a feed, a Trigger Agent applies a rule, an Email Agent tells you. I did not run it for this post, because it is a Rails application with a database and the two above needed one container each; I would reach for it when the job is scraping a page that has no feed, which is where its agents are strongest.
Where the model goes, and where it does not
Look at the two flows again and count the nodes with judgement in them. One. The trigger is a clock. The fetch is a URL. The filter is a word match. The write is a path. Only the node that asks which of these a human needs to see now is a decision I could not write as a rule in advance, and even there my stub rule caught the obvious cases. When I add a model to that node I add exactly one point of cost, one point of latency and one point of being wrong, and I can see it on the canvas.
That is the discipline these tools enforce by their shape. An agent framework lets the model decide everything, including whether to fetch the feed at all. A workflow tool makes you draw the box the model is allowed to be in. For most always-on jobs in a small company, including most of mine, the box is small.
When a workflow beats an agent
- The trigger is a clock or a webhook and the inputs are structured. A feed, a form, a row in a sheet.
- Every step but one can be written as a rule, and you can say which one.
- The job has run for years without a model and only needs one for a single step.
- You want to see the pipeline on a canvas when it breaks at 3am, not read a transcript of what a model decided to do.
And the reverse: when the inputs change faster than you can redraw the boxes, or the decision spans several steps, that is when I move to one of the frameworks in the open-source comparison and accept the transcript.
What I tested, and what I did not
Both flows ran in fresh containers on this Mac, memory-capped, with the model step stubbed as a rule, against a live public feed, and the logs above are from those runs. The first n8n run failed with a 502 from the feed host, which took nine seconds to answer a plain request from my machine at the same moment; the second run, a few minutes later, went through unchanged. That is an ordinary morning for an always-on job, and it is why the timer, not the run, is the unit of reliability. Huginn I described from its README and did not run. None of the three needed anything installed on the Mac itself, which is the other reason they make good first always-on jobs.
Sources
- n8n, n8n README on GitHub (accessed October 20, 2026)
- Huginn, Huginn README on GitHub (accessed October 20, 2026)
- Node-RED, Node-RED repository on GitHub (accessed October 20, 2026)
- Node-RED, About Node-RED (accessed October 20, 2026)
Written by Lera Leonteva. Both container runs are reproduced above; star counts and licence lines are as shown on GitHub on the access date.