Week 43 / 2026.
A short record of changes between issues. Full answers and charts live on the question pages.
- Data window
- Oct 19–25, 2026
- Last updated
- Nov 6, 2026
- Author
- Romuald Członkowski
Commentary
Between 19 and 25 October, n8n-mcp produced 83,032 workflows, 11,862 a day on average, across 1,850,163 tool calls with a 0.12% error rate. That is 1% more workflows a day than the week before and 8% fewer calls, against 163,883 active installs, up 8%; 8% of them create workflows. The busiest day was 21 October with 14,501 workflows, the quietest 24 October with 6,874. Week-one retention stands at 39% of the 6,392 installs that started in the week of 12 October, the same as in the previous issue.
The call mix breaks down as 59% reads, 12% writes, 7% documentation and validation, and 3% tests. Fetching executions alone accounts for 37% of all calls, and every workflow created comes with 7.2 updates. The window holds 143,713 edits: all of them went through the API, 55% produced a verifiable improvement, at an average of 2.69 operations and a median of 880 ms. In this data, working on a workflow is mostly reading and correcting what already stands rather than writing it once. Price it that way.
The window added 29,329 workflows to the corpus, which now holds 1,247,822, with a median of 7 nodes. Code appears in 77% of them, HTTP Request in 62%, Data Table in 12.4%, an LLM model node in 5% and an AI Agent in 3%. The model-node share has been falling since March, when it stood at 17%, while Data Table climbed over the same period from 3.3% to 12.5% in the October data. The week's most searched term is data table, with 202 installs behind it, ahead of webhook (149) and http request (66). Among model nodes, OpenAI holds 32%, Anthropic 21%, Google Gemini 17% and OpenRouter 16%.
Validation turned away 0.7% of workflows, and again it is the smallest ones that trip over it: 1.16% at 2–3 nodes, against 0.09% at 101–300. 13% of connection tests end in an error, and 30% of those come down to an address or the network. This issue's pitfall sits in loops: in a Code → HTTP Request → Code → If arrangement, the HTTP Request node overwrites $json with the API response, so the iteration counter and anything accumulated are lost on every pass. Keep loop state in $getWorkflowStaticData('global') and reset it at the start of every run, because it is saved to the database and carries over between runs.
Changes versus the previous issue
| figure | value | change |
|---|---|---|
| Workflows created | 83,032 | — |
| Workflows per day | 11,862 | +1% |
| Tool calls | 1,850,163 | -8% |
| Calls ending in error | 0.12% | -20% |
| Failed validation | 0.7% | +3% |
| Active installs | 163,883 | +8% |
| Week-1 retention | 38.6% | -0% |
| Workflows added to corpus (week) | 29,329 | +1% |
This week's answers
- Which LLM powers n8n AI agents: OpenAI, Anthropic, Gemini or Ollama?
In n8n workflows built last week, OpenAI holds 32% of model nodes, Anthropic 21%, Google Gemini 17% and OpenRouter 16%. 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 77% of last week's workflows, followed by HTTP Request (62%) 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 Jul 2026 to 3.1% in Oct 2026, a 6.7× increase. Extract from File is second (3.3×).
- Do AI-built n8n workflows handle errors?
Rarely. 2.4% of AI-built n8n workflows from last week contain an Error Trigger node, and 2.4% of those built in Oct 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 77% of AI-built n8n workflows from last week and HTTP Request in 62%. In Oct 2026 the Code share was 76%. 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?
53% of workflows created last week have at least 11 nodes and 19% have 31 or more. The larger the workflow, the less often it fails validation: 1.2% for 2–3 nodes versus 0.16% for 31–100.
- Which n8n nodes do AI agents search for?
Last week the most searched n8n node was data table (202 installs), ahead of webhook (149) and http request (66). The ranking counts unique installs, not queries.
- How do AI agents (Claude Code, Cursor) build n8n workflows through MCP: build or maintain?
Maintain. Of 1,850,163 tool calls last week, 59% read workflows, executions and lists, and 12% write. Fetching executions alone is 37% of calls. In the same week 83,032 workflows were created, 11,862 a day.
Most searched nodes
data table (202) · Webhook (149) · http request (66) · Telegram (49) · whatsapp (46) · anthropic (40) · Slack (35) · Gmail (29) · schedule trigger (28) · gemini (24)
$getWorkflowStaticData global loses data in loops
When looping via Code → HTTP Request → Code → If → back to HTTP Request, the HTTP Request node overwrites $json with the API response. Any loop state passed via $json (iteration counter, accumulated data) is lost on every iteration. Fix: use $getWorkflowStaticData('global') to persist loop state across iterations. Always reset staticData at the start of each workflow run because it persists across runs (saved to DB) — stale data from failed executions can leak in otherwise.