Comparison

flowwork vs n8n

n8n asks you to wire up a canvas. flowwork lets you describe what you already know — your Claude Code agent writes the workflow as code, runs it in your workspace, and you keep an artifact you can actually read, review, and audit.

Both let you automate work. The difference is where the work lives and who does it: n8n is a visual automation server you click together and administer; flowwork is a code-native engine your agent authors and operates, co-located with your files and tools. If you hold the context worth automating and you're already working in Claude Code, that distinction is the whole story.

flowwork n8n
Authoring Code — Claude Code writes a typed, versioned definition with you (the flowwork skill) Visual canvas; Claude can draft node JSON, but the editor is the real authoring surface
Who operates it Your Claude Code agent — authors, triggers, monitors, handles human-in-the-loop You, in the web editor; plus schedules & webhooks
Agent ↔ workflow The workflow can delegate back to the driving agent (the agent directive) LLM & agent nodes call model APIs as steps
Your workspace (files, repo, devops, browser) First-class — the agent operates your local workspace Server-side; reaching your files needs custom / SSH nodes
Claude skills & slash-commands Reusable inside workflows
Source of truth Code — git history, diffs, PR review Workflow JSON in n8n's database (export / import)
Footprint Single binary + CLI on any laptop, VPS, or sandbox Node server + web editor + database (self-host or cloud)
Agent-operable runtime Everything via CLI — author, trigger, inspect, resume; fully scriptable by an agent Browser editor is the control surface; awkward for an agent to drive headless
Ephemeral & sandbox runs Spin up in a sandbox, automate, destroy — pay cents; one Pro license spans up to 5 machines A server you keep running (hosted subscription, or self-host you operate)
Visual builder Read-only graph view in the console Yes — the core experience
Prebuilt integrations Connectors + anything Claude Code can do Hundreds of prebuilt nodes (n8n's strength)
Determinism & resilience Compiled graph: retries, backoff, leases, replay Workflow engine with retries; less strongly typed
Best for Context-rich domain experts & developers automating in Claude Code No-code / low-code teams; broad SaaS glue
Code, not a canvas

A node graph looks simpler — until it isn't.

A visual canvas feels approachable, but it's still a black box you can't diff, review in a pull request, or hand to a teammate with confidence. Once a workflow grows past a handful of nodes, the canvas is the complexity, just hidden.

With flowwork the artifact is code — and because Claude Code writes it with you, writing was never the barrier. You bring the context and the judgement; the agent produces a typed, versioned definition you can actually read, review, audit, and change in seconds.

The inversion

The agent drives the engine — and the engine calls the agent back.

n8n has AI and agent nodes: the workflow calls a model as a step. flowwork inverts that relationship. Your Claude Code agent is the operator that authors and runs the workflow, and a step can delegate back to that agent with the agent directive — to read your files, reason over a document, run a command, or drive a browser.

Your context lives in your files, docs, and repositories. flowwork runs right there in your workspace, so the agent can reach and operate it. n8n executes on a server, away from where your context actually is.

Yes, Claude can write n8n too

Claude Code can draft n8n JSON. Code-first still wins for agents.

We won't pretend otherwise: Claude Code can generate an n8n workflow's node JSON. But authoring is only half the loop — the agent also has to operate the thing. n8n's control surface is a browser editor; an agent running in a terminal or a sandbox can't click it, watch a run, or resume a stuck step without a human stepping in.

flowwork is code plus a CLI the whole way down. The same agent that writes the workflow also triggers it, reads its state, answers human-in-the-loop steps, and ships a new version — no UI in the loop. And because a workflow is just code with a connector-calling convention, the model writes genuinely complex logic — branches, loops, fan-out, retries — as naturally as it writes any function.

Disposable by design

Spin up a sandbox, automate, tear it down — pay cents.

Because flowwork is a single binary the agent drives over a CLI, you can run it inside an ephemeral sandbox: spin one up, let Claude Code author and run the workflow, then destroy it. You pay for the minutes you used — cents — instead of a monthly bill for an always-on automation server.

Licensing is built for this: one Pro license covers up to five concurrent machines, and a sandbox releases its slot the moment it tears down (idle ones expire on their own), so your laptop and a fleet of short-lived sandboxes share a single key. It's the automation throughput of a hosted plan without the hosting — the kind of thing an n8n subscription exists to sell you.

Runs where you do

A CLI you co-locate with the agent.

Drop flowwork next to Claude Code on your laptop, a VPS, or a sandbox and the two operate together — scriptable, disposable, no UI to babysit. Your context lives in your files and repos, and the engine runs right there with the agent.

n8n is self-hostable, but even self-hosted it's a long-running server, a web editor, and a database you administer. That's the right shape for a shared team canvas; it's the wrong shape for an agent working in a sandbox.

When n8n is the better choice

We'd genuinely point you to n8n when:

flowwork is built for the opposite case: you have the context, you're working in Claude Code, and you want durable, auditable automation you own as code.

See it in your own workspace

Free tier: 3 active workflows, 10 runs/day. No credit card.