N8n Architect

n8n Architect

Use this skill for all n8n-as-code work: workspace readiness, environments, managed local instances, tunnels, workflow authoring, validation, sync, push, and pull.

Use npx --yes n8nac as the primary interface. Use npx --yes @n8n-as-code/n8n-manager only for local managed runtime lifecycle, tunnels, and workflow presentation commands that are explicitly exposed by n8n-manager.

Context Root Protocol

• Treat the current context root as the directory containing n8nac-config.json, AGENTS.md, .agents/skills, and the configured workflowsPath.
• Generated context root hint: not embedded. Use the shell launch directory or the workspace path explicitly given by the user.
• Before any n8n work, first run npx --yes n8nac update-ai from the context root, then read AGENTS.md. update-ai is designed to create or refresh the n8n-as-code block without destroying existing user or agent instructions.
• Use the exact n8nac command and n8n-manager command listed in AGENTS.md. Those context-root commands override the portable examples in this skill.
• Run every npx --yes n8nac env ..., npx --yes n8nac workspace ..., npx --yes n8nac list, pull, push, validate, test, and update-ai command from the context root unless the user explicitly gives another context root.
• AGENTS.md is bootstrap context only, not a source of configuration truth.
• Do not infer environment, project, or workflowsPath from AGENTS.md.
• Before n8n work, resolve the effective context from the backend:

npx --yes n8nac env status --json

• Use the returned workflowsPath for workflow files. It is the configured workflow directory for the active environment.
• Do not reconstruct workflowsPath from environment name/id, instanceIdentifier, instanceUserIdentifier, projectId, projectName, or legacy sync fields.
• Never write n8nac-config.json, ~/.n8n-manager, or n8n-manager secret files by hand.

Workspace Readiness

Resolve the effective environment through the backend before workflow work:

npx --yes n8nac env status --json

• Treat env status --json as the source of effective workspace readiness.
• Do not infer readiness from raw files, generated agent docs, or directory names.
• If env status --json fails because the workspace is not configured, use env add, env auth set, and env use to create or select a V4 workspace environment.
• Managed local instances remain machine-global runtime resources.
• Workspace environments remain workspace-scoped and are managed through npx --yes n8nac env ....

Bootstrap Order

1. cd to the context root.
2. Run npx --yes n8nac update-ai, then read AGENTS.md.
3. Run npx --yes n8nac env status --json.
4. If the context root is not ready, inspect managed local instances with npx --yes @n8n-as-code/n8n-manager instance list.
5. Reuse an existing environment or managed local instance when suitable.
6. If no suitable environment exists, stop and ask the user whether they want to connect a remote n8n URL or create/reuse a managed local n8n instance. Do not create infrastructure by default. If the user chooses a managed local instance, ask separately whether they want a public tunnel.
7. Ask for host/API key only for an explicitly remote n8n environment.
8. Configure the environment with:

npx --yes n8nac env add <name> --base-url <url> --workflows-path workflows/<name>
npx --yes n8nac env auth set <name> --api-key-stdin
npx --yes n8nac env use <name>

For a managed local instance:

npx --yes n8nac env add Local --managed-instance <id> --workflows-path workflows/local
npx --yes n8nac env use Local

9. Run npx --yes n8nac update-ai after changing environments when the facade does not do it automatically.

Environments

Use npx --yes n8nac env ... for workspace environments, remote URLs, active environment, API-key binding, projects, and workflow paths.

npx --yes n8nac env status --json
npx --yes n8nac env list
npx --yes n8nac env add <name> --base-url <url> --workflows-path workflows/<name>
npx --yes n8nac env auth set <name> --api-key-stdin
npx --yes n8nac env use <name>

• Prefer --api-key-stdin for API keys.
• Do not pass secrets inline in shell arguments.
• env auth set binds the key to one environment, so several environments may share a base URL with one key each. Run it once per environment; apiKeySource in env status --json is workspace-environment when the environment uses its own key.
• Do not ask for host/API key when the user wants a managed local Docker instance.
• Do not print API keys or credential secret values back to the user.
• If a command or flag is unfamiliar, run npx --yes n8nac env --help or npx --yes n8nac env <subcommand> --help.

Attach a managed local instance to the workspace with npx --yes n8nac env ...:

npx --yes n8nac env add Local --managed-instance <id> --workflows-path workflows/local
npx --yes n8nac env use Local

Managed Local Runtime

Use npx --yes @n8n-as-code/n8n-manager only for local managed instance lifecycle, tunnels, and workflow presentation commands that are part of the local runtime layer.

Inspect existing managed instances before changing local machine state:

npx --yes @n8n-as-code/n8n-manager instance list
npx --yes @n8n-as-code/n8n-manager instance --help
npx --yes @n8n-as-code/n8n-manager config get

Do not invent n8n-manager subcommands. Use npx --yes @n8n-as-code/n8n-manager <subcommand> --help when unsure.

When the context root is not configured and no suitable existing instance is available, stop and ask the user to choose. Do not create infrastructure by default.

Present these choices clearly:

• use an existing managed local instance if one is available;
• create a new managed local n8n instance;
• configure a remote n8n URL as a workspace environment through npx --yes n8nac env.

If the user chooses a managed local Docker instance, ask the tunnel question separately:

• without public tunnel: local n8n only, suitable for normal UI/API workflow work;
• with public tunnel: exposes the instance through a public URL, useful for webhooks/forms/chat triggers and remote callbacks.

Do not enable, refresh, or start a public tunnel unless the user explicitly requested public access, webhook testing, or approved the tunnel option. If public access is not needed, create/start the managed instance without --tunnel.

Only run these commands after the user has explicitly chosen the corresponding option.

Managed local instance without public tunnel:

npx --yes @n8n-as-code/n8n-manager instance create
npx --yes @n8n-as-code/n8n-manager instance start <id>
npx --yes @n8n-as-code/n8n-manager instance list

Managed local instance with public tunnel:

npx --yes @n8n-as-code/n8n-manager instance create
npx --yes @n8n-as-code/n8n-manager instance start <id>
npx --yes @n8n-as-code/n8n-manager tunnel start <id>

Instance and tunnel operations are per managed local instance:

npx --yes @n8n-as-code/n8n-manager instance start <id>
npx --yes @n8n-as-code/n8n-manager instance stop <id>
npx --yes @n8n-as-code/n8n-manager instance remove <id>
npx --yes @n8n-as-code/n8n-manager tunnel start <id>
npx --yes @n8n-as-code/n8n-manager tunnel stop <id>

• Do not delete local instance data unless the user explicitly asks for destructive deletion.
• If Docker is unavailable or the daemon is stopped, report the backend diagnostic and stop. Do not loop.
• If a command fails repeatedly, stop after two attempts and explain the backend diagnostic.

Sync Discipline

• Pull before reading or modifying an existing workflow.
• Push after every modification.
• Use list to inspect workflow IDs, file paths, and sync status.

npx --yes n8nac list
npx --yes n8nac pull <workflowId>
npx --yes n8nac push <path-to-workflow.workflow.ts> --verify

• push requires the full workflow file path, either absolute or context-root-relative. Do not pass a bare filename.
• On n8n 2.x, pushing to a published workflow also releases it to production — the API re-publishes on update. Treat every push to a published workflow as a deploy. Use push --draft when the user wants to check the change in n8n first: it re-pins the previously published version so production keeps running what it already ran.
• For a new workflow, create the file inside the workflowsPath returned by env status --json, then confirm it with npx --yes n8nac list --local.
• If push/pull reports a conflict, use explicit resolution commands. Do not overwrite remote changes blindly.
• pull and conflict resolution operate on a single workflow ID.
• list is the lightweight command that covers all workflows at once.
• If you skip pull, a later push can be rejected by optimistic concurrency control when the remote changed.

Conflict Handling

If push or pull reports a conflict, stop and inspect the conflict. Use explicit resolution commands only after choosing the intended direction:

npx --yes n8nac resolve <workflowId> --mode keep-current
npx --yes n8nac resolve <workflowId> --mode keep-incoming

• keep-current force-pushes the local version.
• keep-incoming force-pulls the remote version.
• Never silently force-push over a remote change.

Schema-First Research

Never guess n8n node parameters.

npx --yes n8nac skills examples search "<workflow pattern>"
npx --yes n8nac skills search "<node or capability>"
npx --yes n8nac skills node-info <nodeName>
npx --yes n8nac skills validate <workflow.workflow.ts>

• Use exact node type and valid typeVersion values from node-info.
• Use exact resource, operation, option, and parameter names from schema output.
• Do not invent parameters, operations, credential types, or CLI flags.
• Treat schema output as the absolute source of truth even if examples or memory disagree.
• Prefer the highest valid typeVersion returned by schema output.
• For fixed collections such as Switch/If rules, Wait form fields, or nested options, read the full node-info output before writing values.

Optional Native n8n MCP Assist

The n8n-as-code MCP server is a client adapter for N8NAC tools. The native n8n MCP server is a separate live n8n instance endpoint. Native n8n MCP can complement this workflow for native knowledge, live state, and runtime execution, but it does not replace npx --yes n8nac, bundled knowledge, .workflow.ts, Git, or the sync discipline.

Use this routing policy:

• Default to local npx --yes n8nac for code-first workflow authoring, validation, pull, push, credentials, execution history, and presentation. Use npx --yes n8nac skills as the bundled offline knowledge default.
• Native MCP assist is configured per n8n-as-code environment. When creating or updating an environment, offer to configure it with npx --yes n8nac native-mcp configure <environment> --token-stdin; do not ask the user to manually configure a separate MCP server for Claude Code or the VS Code Workbench.
• If native MCP assist is configured, use it where it complements n8n-as-code: read-only live discovery, server-side validation, native SDK/reference knowledge, live node definitions, credential metadata without secrets, execution inspection, projects, folders, and explicit runtime execution/test strategy when supported.
• Check native availability with npx --yes n8nac native-mcp status --include-tools --json before relying on native tools.
• For user requests about the current/live n8n instance, existing remote workflows, available nodes in this instance, credential metadata, projects, folders, executions, drift, or duplicate discovery, prefer native MCP read-only tools after the status check. Do not fall back to local npx --yes n8nac list, fetch, verify, or bundled skills as the primary source for those live-audit facts when native MCP read-only tools are available.
• Do not expose native MCP assist on non-loopback HTTP/SSE transports unless the MCP transport is authenticated and N8NAC_NATIVE_MCP_ALLOW_REMOTE=1 is explicitly set.
• Do not request full live execution payloads with includeData=true unless the user explicitly needs payload data and N8NAC_NATIVE_MCP_ALLOW_EXECUTION_DATA=1 is set.
• Prefer npx --yes n8nac test when the execution strategy is to exercise the real webhook, chat, or form trigger contract.
• Prefer native runtime execution only when the generated execution strategy explicitly calls for it and it does something better than npx --yes n8nac test, such as workflow ID execution, non-webhook workflow testing, native pin-data test preparation, or direct execution diagnostics.
• Treat native execute/test as a side-effecting runtime action, like npx --yes n8nac test; do not run it just because the tool exists.
• Do not use native MCP create, update, publish, unpublish, archive, or destructive data-table tools unless the user explicitly requests direct native MCP mode and the tool is gated by permissions.
• If a workflow is ever created or changed through native MCP direct mode, immediately pull it back with npx --yes n8nac pull <workflowId> so the .workflow.ts file and Git remain the source of truth.
• If native MCP validation and local validation disagree, stop and report the divergence instead of forcing a push or direct update.
• Never put native MCP tokens in project files, generated docs, command arguments, or responses.

Use-case routing examples:

• Workflow authoring, editing, pull, push, sync, credentials, and durable workflow changes: use local npx --yes n8nac commands and .workflow.ts files.
• Offline node knowledge, examples, documentation, and schema-first authoring: use local npx --yes n8nac skills commands first.
• Live workflow discovery, drift investigation, projects, folders, credentials metadata, duplicate discovery, and execution inspection: prefer native MCP read-only tools when configured because the user is asking for current instance state.
• Connected-version node definitions or server-side validation: prefer native MCP read-only tools when the user asks what is available in this instance or needs validation against the connected n8n version. Use bundled knowledge for offline authoring when live instance state is not needed.
• Runtime execution: prefer npx --yes n8nac test for real webhook, chat, or form trigger contracts; prefer native runtime execution only for explicit workflow-ID execution, non-webhook testing, native pin-data preparation, or direct execution diagnostics.
• Direct native workflow creation, update, publish, unpublish, archive, or destructive operations: do not use them as an automatic path; require an explicit direct-native request and sync-back plan.

Do not treat the presence of any MCP server as permission to call native n8n MCP tools. Native n8n MCP is used if and only if the generated execution or investigation strategy needs live n8n capabilities that local N8NAC cannot provide as well.

Native MCP assist is a complementary knowledge, live-state, and runtime enrichment path, not the primary authoring or sync path.

Knowledge Commands

Use these commands instead of guessing:

npx --yes n8nac skills search "<node or capability>"
npx --yes n8nac skills node-info <nodeName>
npx --yes n8nac skills node-schema <nodeName>
npx --yes n8nac skills docs "<topic>"
npx --yes n8nac skills guides "<topic>"
npx --yes n8nac skills examples search "<workflow pattern>"
npx --yes n8nac skills examples info <id>
npx --yes n8nac skills examples download <id>

• Start with examples search when the user asks for a common automation pattern.
• Use examples to learn patterns, not as authority over current node schemas.
• If a command or flag is unfamiliar, run npx --yes n8nac <subcommand> --help; do not invent flags.

Workflow Authoring Rules

• Use TypeScript decorators from @n8n-as-code/transformer.
• Regular nodes connect with source.out(0).to(target.in(0)).
• AI sub-nodes connect with .uses(), never .out().to().
• ai_tool and ai_document connections are arrays: ai_tool: [this.Tool.output].
• Other AI connection types are single refs, such as ai_languageModel: this.Model.output.
• They also accept an array when a node exposes several inputs of the same type, where the position is the input index: ai_languageModel: [this.Model.output, this.FallbackModel.output] (fallback model, Model Selector).
• Check node-info for connection-dependent boolean flags before declaring .uses() connections.

Every .workflow.ts file starts with a <workflow-map> block. Read that map first, locate the property name you need, then read only the relevant class section.

Minimal Workflow Structure

import { workflow, node, links } from '@n8n-as-code/transformer';

@workflow({
  name: 'Workflow Name',
  active: false
})
export class MyWorkflow {
  @node({
    name: 'Descriptive Name',
    type: '/* exact type from node-info */',
    version: 4,
    position: [250, 300]
  })
  MyNode = {
    /* parameters from node-info */
  };

  @node({
    name: 'Next Node',
    type: '/* exact type from node-info */',
    version: 3,
    position: [520, 300]
  })
  NextNode = {};

  @links()
  defineRouting() {
    this.MyNode.out(0).to(this.NextNode.in(0));
  }
}

Expression Syntax

• Prefer modern expressions: {{ $json.fieldName }}.
• Use specific-node expressions when needed: {{ $('Node Name').item.json.field }}.
• Avoid legacy $node["Name"].json.field unless you are preserving an existing workflow and have a reason.
• In Switch/If comparisons, value1 is the expression being evaluated and value2 is the literal comparison value.

Node Naming

• Use descriptive names such as Get Customers, Send Slack Alert, or Normalize Payload.
• Avoid names like Node1, HTTP Request, or Code when a more specific name is available.
• Connection references must match the exact node property names in the TypeScript class.

Reading Workflow Files Efficiently

Use the <workflow-map> block as the index before loading large workflow files.

// <workflow-map>
// Workflow : My Workflow
// Nodes   : 12  |  Connections: 14
//
// NODE INDEX
// Property name                    Node type (short)         Flags
// ScheduleTrigger                  scheduleTrigger
// AgentGenerateApplication         agent                      [AI] [creds]
// OpenaiChatModel                  lmChatOpenAi               [creds] [ai_languageModel]
// Memory                           memoryBufferWindow         [ai_memory]
// GithubCheckBranchRef             httpRequest                [onError->out(1)]
//
// ROUTING MAP
// ScheduleTrigger
//   -> Configuration
//     -> BuildProfileSources -> LoopOverProfileSources
//
// AI CONNECTIONS
// AgentGenerateApplication.uses({ ai_languageModel: OpenaiChatModel, ai_memory: Memory })
// </workflow-map>

Navigation rule:

1. Read <workflow-map> first.
2. Locate the property name you need.
3. Search for that property in the file.
4. Read only the relevant node or routing section unless broader context is required.

AI And LangChain Node Rules

AI sub-nodes are not regular data-flow nodes.

@links()
defineRouting() {
  this.ChatTrigger.out(0).to(this.AiAgent.in(0));

  this.AiAgent.uses({
    ai_languageModel: this.OpenaiModel.output,
    ai_memory: this.Memory.output,
    ai_outputParser: this.OutputParser.output,
    ai_tool: [this.SearchTool.output],
  });
}

• Use .uses() for language models, memory, tools, parsers, embeddings, vector stores, retrievers, and other AI sub-nodes.
• Never connect AI sub-nodes with .out().to().
• ai_tool and ai_document must be arrays; every entry lands on input index 0.
• Most other AI connection types are single refs, or an array when the node exposes several inputs of the same type — position = input index.
• needsFallback: true (Agent, Basic LLM Chain) needs a second model on input 1: ai_languageModel: [this.Model.output, this.FallbackModel.output]. Same for the Model Selector node.
• Some nodes require boolean flags to expose AI ports or gated parameters. Check node-info before declaring .uses().

Common Mistakes To Avoid

• Wrong node type: use the exact full type returned by schema output, including package prefix when provided.
• Outdated or non-existent typeVersion: use a value from the schema output.
• Invalid operation/resource value: use exact option values from the schema.
• Mismatched resource and operation: each resource enables its own operations.
• Guessing nested structures: fixed collections have exact shapes.
• Wrong connection names: match TypeScript property names exactly.
• Inventing nodes, credentials, operations, or parameters.
• Connecting AI sub-nodes with .out().to().
• Using ai_tool: this.Tool.output instead of ai_tool: [this.Tool.output].
• Inverting Switch/If value1 and value2.
• Using old Wait form structures such as formFieldsUi.fieldItems when the current schema expects formFields: { values: [...] }.
• Passing a bare filename to push.
• Treating Class A runtime/config gaps as workflow-code bugs.

Verify, Test, And Present

After pushing:

npx --yes n8nac verify <workflowId>
npx --yes n8nac test-plan <workflowId> --json

For webhook, chat, or form workflows, prefer the production test sequence:

npx --yes n8nac workflow activate <workflowId>
npx --yes n8nac test <workflowId> --prod

• Class A configuration gaps require user/config action, not workflow rewrites.
• Runtime-state issues such as unarmed test webhooks are not workflow-code bugs.
• Class B wiring errors are fixable in the workflow file.
• Stop after two repeated failures with the same diagnostic.

Workflow Presentation Contract

npx --yes n8nac workflow present is the standard way to show a workflow to the user. It is v4-environment aware and part of the workflow authoring loop.

Run it whenever one of these is true:

• you created a workflow;
• you modified and pushed a workflow;
• you ran or tested a workflow and the user needs to inspect it;
• the user asks to show, open, present, display, or give the URL/link for a workflow.

npx --yes n8nac workflow present <workflowId> --json

Rules:

• Do not manually construct n8n workflow URLs.
• Do not return an internal local n8n URL when a presentation URL is available.
• Use the url returned by workflow present --json as the user-facing URL.
• If you do not know the workflow ID, run npx --yes n8nac list first and select the matching workflow.
• Do not call npx --yes @n8n-as-code/n8n-manager presentWorkflowResult; it is a legacy runtime command and is not workspace-environment aware.
• If workflow present fails, report the backend diagnostic and then provide the best direct n8n URL only as a fallback.
• Do this before the final response when the task created, changed, pushed, ran, or explicitly asks to show a workflow.

Testability Protocol

For webhook, chat, or form workflows:

1. Push with verification when possible.
2. Run test-plan to inspect trigger type, endpoint, and suggested payload.
3. Activate the workflow.
4. Test with --prod by default.

npx --yes n8nac push <path-to-workflow.workflow.ts> --verify
npx --yes n8nac test-plan <workflowId> --json
npx --yes n8nac workflow activate <workflowId>
npx --yes n8nac test <workflowId> --prod

Use bare npx --yes n8nac test <workflowId> only when a test URL was intentionally armed in the n8n editor.

For GET/HEAD webhooks that read from $json.query, prefer:

npx --yes n8nac test <workflowId> --query '{"key":"value"}' --prod

Execution Debugging

If a webhook returns success but the workflow behavior is wrong, inspect executions instead of guessing:

npx --yes n8nac execution list --workflow-id <workflowId> --limit 5 --json
npx --yes n8nac execution get <executionId> --include-data --json

• A successful HTTP trigger only means n8n accepted the request.
• The execution can still fail later inside the workflow.
• Use execution data to identify the failing node and real payload shape.

Credential Workflow

When a workflow is blocked by missing credentials, resolve the credential gap without rewriting unrelated workflow logic.

npx --yes n8nac workflow credential-required <workflowId> --json
npx --yes n8nac credential schema <type>
npx --yes n8nac credential list --json
npx --yes n8nac credential create --type <type> --name <name> --file cred.json --json
npx --yes n8nac workflow activate <workflowId>

• workflow credential-required exits non-zero when at least one credential is missing. Treat that as a signal to act, not as a workflow-code failure.
• Use credential schema to discover required fields.
• Ask the user for secret values when needed.
• Prefer --file for credential creation. Do not pass secrets inline in shell arguments.
• Do not print API keys or credential secret values back to the user.
• If credential creation fails, read the validation message and change the payload before retrying.

Operating Loop

For most workflow tasks:

1. Resolve context with env status --json.
2. Read workflowsPath from the backend response.
3. Inspect existing workflows with list.
4. Pull before editing an existing workflow.
5. Search examples and schemas.
6. Edit or create the .workflow.ts file.
7. Validate locally.
8. Push with --verify.
9. Test if the workflow is HTTP-triggered.
10. Inspect executions when behavior is unclear.
11. Present the final workflow link with npx --yes n8nac workflow present <workflowId> --json.

Response Discipline

• Explain concrete actions and command results, not generic capability.
• When the user asks for an URL or visual inspection of a workflow, run npx --yes n8nac workflow present <workflowId> --json instead of composing a URL manually.
• If setup is missing, use n8nac env ... for workspace environments and n8n-manager only for managed local instances.
• Do not ask for host/API key unless the user chooses a remote n8n environment.
• Do not tell the user to run setup commands when you can run non-interactive commands yourself.
• Stop after two repeated failures with the same diagnostic and report the backend error clearly.