Source: wiki synthesis: n8n vs Claude Code — When to Use Which, Building n8n Workflows with Claude Code, The Loop Is the Unit of Work, Should You Build a Loop? · 2026-09-29 addition: raw/Build_Sell_with_Codex_5+_Hour_Course.md (Nate Herk, “Build & Sell with Codex” 5+ hour course, YouTube)
Every automation task hits the same fork: wire it as a deterministic visual workflow, or hand it to an agent loop. The n8n topic answers from the workflow side (Nate Herk’s decision framework), the agent-loops topic from the loop side (the four-condition test, the token math, the security tax) — but neither cluster reads the other. Composed, they yield a three-way triage in which “keep it a manual prompt” is a first-class outcome, not a failure.^[inferred — the composition is this article’s synthesis] This decision comes before Running the Agentic Loop (which picks where a loop runs once you’ve decided to build one) and is orthogonal to Vendor-Direct Tool Calls (which picks the tool-call surface, CLI/MCP/middleware, for a given vendor action).
Key Takeaways
- Task shape is the first cut. Fixed steps with one right answer — schedule triggers, “if new row in Sheet then send Slack” — are wiring territory: Herk’s 10-minute rule says if it fits a small handful of nodes, just build it in n8n. Open-ended work — branching strategy, “search until I have enough info, then synthesize,” the ~15-node YouTube-dedupe-summarize example that becomes one paragraph in Claude Code — is loop territory.
- Loop territory has an entrance exam. The four-condition test: the task repeats at least weekly, verification is automated (a test/build/linter that can fail the work without you in the room), your token budget can absorb the waste, and the agent has a senior engineer’s tools. “Miss one box, keep it a manual prompt” — a loop pointed at the wrong task “costs more than it returns, forever.”
- Verification is the hidden variable separating the two.^[inferred] In a wired workflow the node graph is the spec — determinism substitutes for verification, and n8n’s per-node execution log shows exactly which step failed. In a loop the path is open-ended, so an objective gate must define “done” instead; without one you get the named failure family — Ralph Wiggum loops, agentic laziness, goal drift, self-preferential grading.
- The economics run in opposite directions. A published n8n workflow runs forever on a 47,000). Counter-current: n8n cloud pricing scales with executions while token prices fall, so high-volume agent workloads trend cheaper in code. The deciding metric for the loop side is cost per accepted change.
- The security tax differs in kind. Loops: community skills as injection vectors, secrets scattered into debug logs, permission scope creep — a 30-day re-audit checklist. Wired workflows carry operational risk instead: a 60-node agent-authored oneshot throttled the n8n cloud API and bricked the instance for over an hour (self-hosted recovers via reboot; cloud needs a support ticket).
- The bridge dissolves the either/or for authoring. As of Q2 2026 n8n ships a first-party MCP server in every edition: Claude Code describes the workflow, and a self-correcting build loop (generate TypeScript that must compile → validate → fix → execute → retry) writes it into your instance. You get agent authoring speed plus workflow deployment determinism — Herk’s framing: “n8n is not dead, it has become the foundation, and Claude Code is the new top of the stack.”
The combined triage
Walk it top-down; stop at the first match.^[inferred — composes Herk’s decision rules with the four-condition test]
- Deterministic and small (10-minute rule), owned long-term by a non-coder, dependent on the 1,900+ integration catalog, or already running in n8n? → Wire it. Publish in n8n and leave it alone — “don’t migrate working automations just because a new tool exists.”
- Open-ended or decision-making, AND all four conditions pass (weekly recurrence, automated gate, budget headroom, real tooling)? → Loop it. Then, and only then, pick the runtime — in-process, durable, or hosted, per Running the Agentic Loop — and write the verifier before the first run per Verifier-First Loops.
- Open-ended but a condition fails — it’s a one-off, there’s no automatable check, the budget can’t absorb waste, or the agent can’t run what it writes? → Keep it a manual prompt. This is “the cheapest decision in the whole topic”: you stay the feedback loop until the task earns automation.
- Complex logic, but you want n8n’s deployment story? → Hybrid. Author the n8n workflow with Claude Code through the MCP bridge; deploy on the canvas where ops can read it.
Build order if branch 2 fires, from the loop-economics source: one manual run that works → a skill → an automation → a state file → a gate → then schedule it. “Skip ahead and you’re paying for a swarm before you have a single run that works.”
What each side charges you
| Dimension | Wired workflow (n8n) | Agent loop (Claude Code + scheduler) |
|---|---|---|
| Marginal cost | $3-8/mo self-hosted VPS, runs forever; cloud scales with executions | 50k-200k tokens per medium single-agent run (author estimate); context re-billed quadratically; a daily-scheduled loop burns millions of tokens a week |
| Where blowups happen | cloud API throttling on huge agent-authored builds (60-node oneshot, >1 hour outage) | context resubmission (50k-300k tokens on a single prompt), tool-output bloat (40-60% removable waste), uncapped subagent fan-out (15k single run; $47k 3-day incident) |
| Observability | per-node input/output execution log built in — “the dashboard tells you which step failed and why” | build it separately (Trigger.dev, Sentry, custom logging) |
| Security surface | operational: cloud recovery requires a support ticket | skill injection, secret scatter in verbose logs, permission scope creep — re-run the 30-day checklist |
| ”Done” is defined by | the node graph — one right answer per shape | an objective gate: pass/fail test, build, zero/non-zero linter — “not a verifier with an opinion” |
| Who can own it | any non-coder — the canvas is legible to ops | a loop engineer who “writes the VISION.md, the gate, and the stop condition, then walks away and trusts the verifier” |
The bridge — hybrid patterns the sources document
-
Claude Code authors n8n. The official first-party MCP server (Public Preview, n8n 2.18.4+, all editions including free self-hosted) builds and updates workflows from a prompt; n8n’s own testing calls Claude Code its strongest client. Known rough edges: complex branching often needs a second pass, and node selection can pick the wrong near-duplicate node.
-
Split by interface. Customer- and ops-facing workflows live in n8n (visual, debuggable by ops staff); custom agent logic, research tasks, and anything that must think for itself lives in Claude Code + a runtime like Trigger.dev.
-
Claude Code as teacher. Keep production automations in n8n; use Claude Code to explain patterns, prototype ideas, and walk new frameworks before you build them for real on the canvas.
-
The moat transfers. n8n fluency — triggers, data flow, error handling, idempotency, observability — is the mental model that makes you good at directing loops; beginners who skip it hit context drift, hallucinations, and scoping walls “without knowing why.”
-
The ladder connects them.^[inferred] The agent-loops leverage ladder (prompt → harness → loop) applies to workflow authoring itself: dragging nodes is the manual rung, prompting Claude Code to build the workflow is the agent rung, and the bridge’s self-correcting build loop (validate → fix → re-validate → execute → retry) is a bounded loop whose verifier — the TypeScript compile + n8n validation — is supplied by the platform.
-
Build with the agent, host as code (Nate Herk, 2026-09-29). In his “Build & Sell with Codex” course (course notes,
raw/Build_Sell_with_Codex_5+_Hour_Course.md), Herk builds automations inside Codex, pushes them to GitHub and hosts them on trigger.dev. His reason: Codex scheduled tasks run as messages in a chat thread and “eat” the weekly subscription limit. He first sorts each automation as deterministic (fixed steps) or non-deterministic, then uses one of three shapes:- Scheduled. A 6am weekday calendar brief sent as a ClickUp DM. AI is used only for the research decision and the draft; the test run cost about 1.33 cents against a 25-cent per-run cap.
- Webhook. A form submission becomes a ClickUp DM with an AI-drafted outreach recommendation and email. He notes the notification alone needs no AI; placeholders would do.
- Codex SDK. A full agent loop, for work where the number of steps isn’t known in advance and has to run at scale outside the subscription. His example, which he currently runs as a Codex scheduled task, is an agent that wakes every 30 minutes, checks a trading account, does research and reads its own past logs so each run isn’t stateless.
His warning for the loop side: an earlier simple agent routine, “after about a month, it started to just go rogue”, posting to the wrong DMs and the team channel “because it was an agent on the back end … interpreting the message different every time.” The fix was a script with the ClickUp channel ID hard-coded, so it “doesn’t have the option to act differently.” He estimates that “90%” of business and personal-productivity automations are the deterministic kind. This is branch 1 of the triage above in practice: once the path is fixed, the agent belongs at authoring time, not at run time. ^[inferred — the mapping to branch 1 is this article’s]
Try It
- Run the triage on your next automation idea and write down which branch fired and why — the one-line record is what stops relitigating the choice later.
- Branch 2? Pre-flight before tokens: name the done-condition, the per-pass check, the saved artifact, and the failure-retry path; instrument cost-per-accepted-change from run one (“tokens-spent and tasks-attempted are vanity metrics”).
- Branch 4? Wire the bridge: enable the official MCP server from n8n instance settings (2.18.4+), connect Claude Code per n8n’s setup guide, and build in plan mode so it asks the clarifying questions first.
- The moment any loop goes unattended, put the 30-day security checklist (SAST + dependency audit + secret scanning in the gate, no skill auto-install, verbose logging off, permissions re-audited, human gate on merge/deploy) on a recurring calendar reminder.
- Prefer self-hosted n8n when agents author workflows at scale — the cloud-throttling incident’s recovery asymmetry (reboot vs support ticket) is the argument.
- Audit your running agent routines for fixed paths. For each scheduled agent task, mark each step deterministic or AI. If only one or two steps need AI, have the agent write a script with hard-coded destinations and a per-run cost cap, and host it outside your subscription (Herk uses trigger.dev), keeping the AI call only where the output changes each run.
Related
- n8n vs Claude Code — When to Use Which — the workflow-side decision framework this article extends with the loop-side tests.
- Should You Build a Loop? — the four-condition test, the quadratic cost math, and the security tax.
- Building n8n Workflows with Claude Code — the hybrid authoring path, including the official first-party MCP server.
- The Loop Is the Unit of Work — the leverage ladder and the
/loop-/goal-Routines-Dynamic-Workflows comparison for what “loop it” signs you up for. - Running the Agentic Loop — the next decision after “loop it”: in-process, durable, or hosted.
- Vendor-Direct Tool Calls — the orthogonal surface decision: CLI vs MCP vs middleware for a given vendor action.
- Verifier-First Loops — the pre-flight checklist branch 2 requires before the first run.
- Write Loops, Not Prompts — the topic entry point whose starter loops are the canonical branch-2 candidates.
Open Questions
- No published benchmark implements the same automation both ways and compares n8n execution cost against loop token cost end-to-end — the comparison table above splices each side’s own numbers.
- The loop-side token ranges are one author’s estimates (50k-200k per medium run), corroborated in order of magnitude but not validated by the 2026-07-03 benchmark drain; the three named starter loops remain individually unmeasured.
- The official n8n MCP server is still Public Preview with no GA date and no feature-parity comparison against the community server — both flagged in the bridge article.