n8n MCP with Claude: build workflows from competitor and AI answer data

The n8n connector is the workflow side. Its Agent Index listing shows 27 tools: 13 reads, 10 writes, 1 admin and 3 it does not classify. Claude can search workflows and executions, look up node types, draft a workflow with create_workflow_from_code, validate and test it, publish it, run it and keep rows in data tables. It has no view of what competitors shipped this week, and it cannot see what the answer engines say about your category.

Linkeddit is the data side. Compete tracks competitor moves and gives each one a priority, a category and a reason it matters. Answer Radar records which brands the five engines name and recommend for your buying questions. Both come with one plan at $60 a month. Linkeddit writes no workflow and sends nothing, so the customer's own assistant decides what to build from the data.

Claude sits between them. It reads the Linkeddit data, builds or updates the n8n workflow, tests it, and stops before publishing so a person reads it first.

n8n's MCP page gives the steps. Turn on Enable MCP access under Settings and Instance-level MCP, copy the Server URL from the OAuth tab, add n8n in Claude's connector list and authorize it. n8n lists version 2.18.3 or higher and says it works on all plans, including the free Community version. The Linkeddit connector is added the same way, at https://mcp.linkeddit.com/mcp.

Two settings matter more than the rest. n8n lets you choose which workflows are exposed to MCP, so Claude only sees what you made available. And you can require Claude to ask permission before its read and write tools run. Set both before the first prompt.

On Linkeddit's own test project, plausible.io, one read-only call to compete_signals_get returned 19 competitor signals for the week of 2026-10-26: 2 high, 9 watch and 8 low priority. The two high ones were a Matomo security update and Fathom's launch of an AI assistant. compete_monitors_get showed 5 monitors, all set to weekly, each with a next run of 2026-11-04.

answer_runs_get on the same project returned three audit runs from 2026-10-27 and 2026-10-28, all with status partial, and the competitor brief was empty. A workflow built on this data should say that. Claude is told to name partial runs and leave out an empty brief, so the report reads as an incomplete week, not a clean one.

The first is a competitor monitoring workflow: a data table of signal ids already seen, a Switch on priority, and an alert for high signals. Unlike the n8n templates that scrape a page and compare it with yesterday's copy, it starts from graded moves that already carry a reason and a source link. The second is a weekly SEO report: the Search Console email n8n's templates already send, plus a section on where the AI engines name a competitor instead of you.

The third is a check that a published change worked. After you publish a fix, Claude estimates a re-check with answer_verify_start, which spends credits, so Linkeddit returns an estimate first and Claude waits for your yes. It later reads the outcome with answer_verify_get and writes a line into the workflow's data table. Each guide has an exact prompt, the steps and a measured read-only run.

n8n should own anything that recurs without you. n8n's own article on Claude Code and n8n says recurring, unattended processes that need retries, logs and version history belong in n8n, and that judgment-heavy steps belong in an AI step inside it. The same article recommends connecting Claude Code to n8n's MCP server to create, edit and test workflows.

So Claude does not run on a timer. It builds the workflow, and it writes the week's Linkeddit data into a data table when you start it. A published workflow can then route and deliver on its own schedule. Linkeddit keeps its own clock too, with competitor monitors that run weekly. n8n also documents an MCP Client Tool node that lets a workflow call an external MCP server, but this site has not tested it against Linkeddit.

Read the workflow Claude drafted, and read the test_workflow result on pinned data. Confirm which workflows are exposed to MCP and that approval is required before write tools run. n8n documents a default five minute MCP execution timeout, and execute_workflow does not support multi-step forms or human steps, so keep the alert path short.

A startup founder wrote on Reddit that many companies "still rely on spreadsheets, saved tabs, screenshots, and occasional alerts for competitor tracking." A workflow with a data table, a priority switch and a named channel replaces that, but only if someone checks that it alerts on the right things first.