n8n Automation MCP
Connect your self-hosted n8n instance to Claude Code. Search real templates and node schemas offline, deploy workflow JSON to your server, then run and auto-repair failures — without a cloud automation host in the middle.
Discovery and draft
- Search nodes with includeOperations=true when configuring integration nodes; use nodeCategoryName or includeBrowseCategories=true to narrow the catalog.
- Search templates with mode=by_metadata (categoryName, minViews), facetField/facetValue, or keyword search; use n8n_index_get_template mode=structure to inspect graph shape before pulling full JSON.
- Use n8n_index_get_ai_patterns mode=detail for node sequences mined from popular templates.
- Assemble workflow JSON with exact node types, names, connections, and parameters.
- Validate offline for graph shape, unknown node types, and connection targets before touching the instance.
Self-healing build loop
When building or repairing a workflow, follow this loop. Retry at most three times before stopping and reporting blockers.
- Draft offline — finalize nodes, connections, and parameters; run offline validation.
- Deploy — create or update the workflow on the instance (full replace or incremental patch via connectBranch / connectCase for IF and Switch wiring).
- Validate live — check instance readiness: triggers present, credential references bound, activation blockers surfaced.
- Run and repair — prefer n8n_run_and_repair for webhook or form workflows: it executes, waits for the correlated execution, applies offline autofix plus execution-aware runtime repairs between attempts, and returns structured status including missingCredentials blockers (set raiseOnFailure=true only when you want a hard error).
- On failure — inspect per-node execution input/output with n8n_get_node_run_data (workflowId, outputIndex, detail=schema), preview autofix on the workflow JSON, apply fixes via n8n_patch_workflow (validateOnly=true to dry-run ops), redeploy or patch, then return to step 3.
Before a retry after a failed run, inject pin data on trigger or credential nodes when you need to simulate upstream payloads without calling external services.
Branching and expression gotchas
Apply these during draft, validation, and autofix — they cause silent failures when wrong.
- IF node: main output index 0 is the true branch, index 1 is false.
- Split In Batches: wire the loop body to output index 1; wire completion or aggregation to index 0.
- Switch: wire each rule to its output index; fallback/default handlers usually belong on the final output.
- Patch wiring: use connectBranch / connectCase (or branch / case on addConnection) instead of guessing output indices.
- Expressions: dynamic values use ={{ ... }} when referencing $json, $node, or other nodes.
- Webhook paths: must be unique per workflow on the instance.
- Credentials: nodes that declare credential slots need bound credential references before activation.
Operational notes
- Catalog search and offline validation need no n8n login; instance actions require a connected n8n server with API access.
- Webhook execution requires an active workflow with a configured webhook trigger.
- Schedule-triggered workflows run on their schedule; they are not manually executable through the webhook path.
