Most of the set is deterministic and offline, so an agent can generate a workflow, lint it and diff it before any credentials exist. The tools that reach a running instance are the exception, and need its URL and API key. A created workflow arrives inactive and needs a second call to switch on, and the diagnosis tools read an execution payload that execution_list can hand over with includeData set to true.
An n8n server that generates, lints and diffs workflow JSON offline, and separately talks to a running n8n instance over its API.
- Workflow JSON generated from a plain-English description, covering webhook, schedule and RSS triggers, common action nodes, and LangChain AI Agent setups wired through ai_languageModel, ai_memory and ai_tool connections.
- A linter that names concrete faults: missing credentials, deprecated node types such as Function against Code, broken connections, missing or non-numeric typeVersion, duplicate node names or IDs, an AI Agent lacking its ai_languageModel sub-node, a Webhook without a webhookId, and an IF node still on the v1 condition schema.
- A semantic diff between two workflows, reporting nodes added, removed or modified with field-level deltas, connection topology changes and settings drift, while ignoring small position moves and timestamps.
- Execution diagnosis from a pasted payload: a per-node summary of nodes that returned 0 items, unresolved expressions, errors with hints and LLM token usage, alongside a timeline of start offset, duration and items in and out.
- A replay workflow that isolates a single node behind a Manual Trigger and a Replay Seed Code node with pinned items.
- Live-instance operations: list workflows with id, name, active, nodeCount, updatedAt and tags, fetch one by id, create one, activate or deactivate it, and list executions by workflow or status.
- A TypeScript skeleton for a custom node implementing INodeType, with a description, a credentials reference and an execute method stub.
N8N_API_URL and N8N_API_KEY for every tool that reaches a live n8n instance; generation, linting, diffing and execution analysis run without them. The diagnosis tools expect the execution JSON pasted in, which the n8n UI's Show details view or the executions API returns.
One command plus a key — npm install -g @automatelab/n8n-mcp, then supply credentials