Migrate from n8n

Convert an n8n workflow export, URL or gallery template into a routine pack. Supported nodes, expressions, code nodes and what needs a person.

routineos migrate n8n <file-or-https-url> --out DIR reads an n8n workflow export (a bare workflow, or a template response that holds it under workflow) and writes a pack and a report. See the overview and the concept map for the full node list. Facts about n8n below carry a source. n8n’s own export format is documented at https://docs.n8n.io/workflows/export-import/.

Supported nodes

n8n	Concept	Becomes
Schedule Trigger	trigger.schedule	the routine schedule, see below
Webhook, Form Trigger, chat triggers, Manual, Execute Workflow Trigger	trigger.webhook, form, chat, manual, subworkflow	the default schedule `0 22 * * 1-5` plus the Desk action queue, with an issue
app triggers	trigger.app	a scheduled poll every hour on the hour
Error Trigger, error workflow setting	trigger.error, failure.workflow	failure handling: an Inbox alert
a node for a known app (Gmail, Slack, Sheets and others)	step.app	a model stage using a connector, read or write
HTTP Request	step.http	a model stage using an `http` connector; headers and auth are never copied
Code, Function	step.code	a code stage with a stub and the original source kept
AI Agent	step.agent	a skill stage with a generated `skills/<id>/SKILL.md`
LLM, summarisation, extractor, classifier chains	step.llm	a skill stage
IF, Switch, Filter	step.branch	a model stage that states the condition
Merge	step.merge	a model stage that states how inputs combine
Loop Over Items	step.loop	a model stage that runs the body for each item
Set, Aggregate, Split Out, Sort, Limit, Rename	step.transform	a model stage
No Operation	step.noop	dropped, recorded as info
Wait	step.wait-time, step.wait-approval	a delay becomes “on the next run”; an approval wait becomes a gate stage
Execute Workflow	step.subworkflow	a model stage naming the child, needs a person
Respond to Webhook	step.respond	output goes to the Desk Inbox, needs a person
Sticky Note	step.note	text goes to the pack readme
Error output (a node set to continue using the error output)	failure.node	steps reachable only from the error output are kept, start with “Only if an earlier step failed:” and raise a needs-person issue; n8n appends the error output after a node’s own outputs (https://raw.githubusercontent.com/n8n-io/n8n/master/packages/workflow/src/node-helpers.ts, fetched 2026-10-30); the slot is verified for IF (2) and one-output nodes (1), unverified for Switch and for Filter’s discarded output
anything else	step.unknown	a “Needs a person” model stage and an `unmapped` issue

Static data between runs becomes a store collection. Per-node error settings (onError, continue on fail) become the stage’s failure rule.

Schedule

The Schedule Trigger parameter shape is read from https://raw.githubusercontent.com/n8n-io/n8n/master/packages/nodes-base/nodes/Schedule/ScheduleTrigger.node.ts (fetched 2026-10-30): rule.interval[] items with a field (seconds, minutes, hours, days, weeks, months or cronExpression), per-field interval keys and time keys (triggerAtHour, triggerAtMinute, triggerAtDay where 0 is Sunday, triggerAtDayOfMonth).

n8n	cron
seconds, minutes	`0 * * * *`, plus an issue: RoutineOS runs at most hourly
hours (N)	`M /N * ` (N=1 gives `M * * *`)
days (N)	`M H /N ` (N=1 gives `M H * *`)
weeks	`M H * * D1,D2`; a week interval above 1 is reported
months (N)	`M H DOM /N `
cronExpression	a 6-field expression drops the seconds field; a minute field that is not one number becomes its first number, or 0, plus an issue

Several rules: the first is used and the rest are reported. The workflow timezone is kept in the plan and reported, because RoutineOS schedules are UTC. The converter does not shift hours. The legacy cron node is treated as unverified and gets the default schedule.

Expressions

A deterministic pattern pass runs over strings that start with = or contain {{ }}. Expression syntax is documented at https://docs.n8n.io/data/expressions/. That page does not cover every variable, so the rewriter is our own and anything it does not recognise is kept and flagged.

Pattern	Rewritten as
`$json.a.b`, `$json["a"]`, `$input.item.json.a`, `$input.first().json.a`	“`a.b` from the previous step”
`$('Node').item.json.a`, `$('Node').first().json.a`, `$node["Node"].json.a`	“`a` from step `<id of Node>`”
`$now`, `$today`	“the run’s start time”
`$execution.resumeUrl`	“the gate decision”, plus an issue
anything else in `{{ }}`	kept verbatim in backticks plus a `needs-person` expression issue

Each rewritten reference is added to the step’s inputs, so stage inputs are explicit. The skill rewrites flagged expressions by hand.

AI nodes

AI Agent and chain nodes become skill stages; their system message and prompt go to the generated skill with expressions rewritten (see https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent/ for the node). Sub-nodes are read from the ai_* connections:

chat model: the model is Claude, and the provider is recorded (step.ai-model)
memory: a store collection memory/<id> (step.ai-memory)
tools: connectors on the owning skill stage (step.ai-tool)
output parser: the output shape is described in the skill (step.ai-output-parser)
vector stores, embeddings, document loaders and text splitters: reported only, because RoutineOS has no vector store (step.ai-vector, unmapped)

Code nodes

The converter writes lib/<id>.py, a stub that prints “not translated yet” to stderr and exits 2, and keeps the original in lib/source/<id>.<lang>.txt. The Code node runs JavaScript or Python (https://docs.n8n.io/code/code-node/). Translating it is the skill’s job, with a person reviewing, or the step is replaced by a model stage. The converter never runs it.

Template search and fetch

migrate search "<words>" [--rows N] calls https://api.n8n.io/api/templates/search and prints id, name and views, plus the gallery total.
migrate fetch <id> --out DIR calls https://api.n8n.io/api/templates/workflows/{id} and converts .workflow. The id must be digits only.
Both use the vetted pack reader: https only, no redirects, public addresses only, a 20 second timeout, and a size cap (1 MiB for search, 2 MiB for a workflow).
The API shape (search returns totalWorkflows, workflows[], filters; one template returns the importable JSON under workflow) was observed on 2026-10-30 by direct request. n8n publishes no documentation or stability guarantee for it that we found, so treat it as undocumented and subject to change. Rate limits: unverified.
Nothing is stored. A fetched template is marked untrusted: generated skills get a third-party banner and the report gets a “Read before you install” section. See limits for the licence question.

Wrapping for a later tool

The functions return plain data, so a later MCP tool can wrap them without the CLI. In migrate_n8n.py:

Function	Returns	Tool
`search(query, rows, opener)`	`{'total', 'results': [{'id', 'name', 'views'}]}`	`search_n8n`
`fetch(template_id, opener)`	`(workflow dict, source dict)`	`convert_n8n` (first half)
`parse(workflow, source)`	a `routineos.migration.v1` plan dict	`convert_n8n` (second half)
`rewrite(text, names_to_ids)`	the rewritten text, with inputs and issues	internal

And in migrate.py: emit(plan) returns (pack dict, MIGRATION.md text), validate_plan(plan) returns a list of problems, and write_out(pack, report, out_dir) writes the two files. All raise migrate.MigrateError with a plain message.