Week 40 / 2026.
A short record of changes between issues. Full answers and charts live on the question pages.
- Data window
- Sep 28, 2026 – Oct 4, 2026
- Last updated
- Oct 16, 2026
- Author
- Romuald Członkowski
Commentary
Between 28 September and 4 October, n8n-mcp produced 73,123 workflows, 10,446 a day on average, across 1,684,118 tool calls with a 0.13% error rate. This is the first issue after I moved telemetry to new infrastructure on 27 September. Installs running versions older than 2.90.0 stopped reporting that day, so I don't hold these figures against earlier issues. There are 143,073 active installs, 8% of them create workflows, and 275 installs hold 31 workflows or more, 2.4% of the creators.
On weekdays the pace stays between 10,557 and 12,192 workflows; the window's two lowest days are its Saturday and Sunday, at 6,962 and 7,477. Validation rejects 0.8% of workflows, and the smaller the workflow, the more often it fails: 1.47% at two or three nodes, 0.62% at four to ten, 0.30% at 11 to 30, and only 0.07% at 101 to 300. The daily rate swings between 0.4% and 3%, so I don't read a single week as a trend, and the data doesn't show what sits behind the spread.
Reads make up 56% of tool calls and writes 12%, while fetching executions alone takes 30% of everything. The window holds 128,924 workflow edits: 2.64 operations each on average, a median of 832 ms, a verifiable improvement in 57% of them. Changing configuration is the most common intent among the labelled edits (48,206), adding functionality is more than twice as rare (18,780), fixing validation accounts for 4,342. Each workflow created draws 7.5 updates, so tuning, not the first launch, is the bigger line in an implementation budget.
In this window's corpus (25,911 workflows, median 7 nodes) Code appears in 75%, HTTP Request in 61%, an LLM model node in 6%, an AI Agent in 4%. Data Table is in 11.9% of workflows and has been climbing since March, when it stood at 3.3%. Among model nodes, OpenAI holds 42%, Anthropic 18%, Google Gemini 15% and OpenRouter 13%. Connection tests fail in 17% of cases, most of them on an address or network error, which is 14% of all tests.
Changes versus the previous issue
On 27 September 2026 I moved n8n-mcp telemetry to new infrastructure. Versions older than 2.90.0 stopped reporting that day, so figures from before and after it come from different sets of installs. I don't compare one with the other.
| figure | value | change |
|---|---|---|
| Workflows created | 73,123 | — |
| Workflows per day | 10,446 | — |
| Tool calls | 1,684,118 | — |
| Calls ending in error | 0.13% | — |
| Failed validation | 0.8% | — |
| Active installs | 143,073 | — |
| Week-1 retention | n/a | — |
| Workflows added to corpus (week) | 25,911 | — |
This week's answers
- Which LLM powers n8n AI agents: OpenAI, Anthropic, Gemini or Ollama?
In n8n workflows built last week, OpenAI holds 42% of model nodes, Anthropic 18%, Google Gemini 15% and OpenRouter 13%. In Mar 2026 OpenAI had 48% and Anthropic 12%.
- Which n8n nodes do AI agents use most, and which trigger nodes start their workflows?
The most used node in AI-built n8n workflows is Code, present in 75% of last week's workflows, followed by HTTP Request (61%) and Webhook (50%). The median workflow has 7 nodes.
- Which n8n node is growing fastest?
The fastest-growing n8n node is stopAndError: its share of AI-built workflows rose from 0.5% in Jun 2026 to 1.7% in Sep 2026, a 3.4× increase. Extract from File is second (2.9×).
- Do AI-built n8n workflows handle errors?
Rarely. 2.5% of AI-built n8n workflows from last week contain an Error Trigger node, and 2.0% of those built in Sep 2026. Since March the share has never exceeded 2.5%.
- Do AI agents build n8n workflows from nodes or from code?
From code. The Code node appears in 75% of AI-built n8n workflows from last week and HTTP Request in 61%. In Sep 2026 the Code share was 73%. Agents would rather write JavaScript and call an API than look for a ready-made integration.
- How large are AI-built n8n workflows, and which fail validation?
52% of workflows created last week have at least 11 nodes and 18% have 31 or more. The larger the workflow, the less often it fails validation: 1.5% for 2–3 nodes versus 0.14% for 31–100.
- Which n8n nodes do AI agents search for?
Last week the most searched n8n node was data table (232 installs), ahead of webhook (150) and http request (103). The ranking counts unique installs, not queries.
- How do AI agents (Claude Code, Cursor) build n8n workflows through MCP: build or maintain?
Maintain. Of 1,684,118 tool calls last week, 56% read workflows, executions and lists, and 12% write. Fetching executions alone is 30% of calls. In the same week 73,123 workflows were created, 10,446 a day.
Most searched nodes
data table (232) · Webhook (150) · http request (103) · schedule trigger (65) · Telegram (57) · Gmail (53) · google sheets (53) · Slack (53) · anthropic (38) · whatsapp (38)
Per-item execution on large datasets
Google Sheets node processes items one at a time by default, making API calls for each row. For large datasets (100+ rows), this is very slow and may hit rate limits. Use batch operations or the "Append or Update" operation with multiple rows when possible.