n8n Review (August 2026): Is the 2.35 Agent Builder Worth the Upgrade?

n8n shipped nine minor versions in the seven weeks since our July review, and three of them decide whether you need a DevOps hire to run production agents. The current stable release is 2.35.5, out August 20, and it is the version this review is pinned to. The 4.5 from July holds. What moved is not the rating but the reason for it: in July the case was self-hosting economics plus a novel MCP server, and the agent node was a bare loop you had to armor yourself. At 2.35 the agent side gained a test harness, per-tool human review, and cost visibility before you run. Those lower the breakage risk on the exact thing the July review flagged as the candid con.

Here is the short version. If you can run Docker, self-host the free Community Edition and skip the meter entirely. If you cannot, the 50 EUR Pro plan is the ops-outsourcing option once you clear 2,500 executions a month. Everything that shipped in 2.35 is real and worth turning on, but every headline feature is a v1 or a Preview, so treat it as production-adjacent, not production-hardened. One platform still beats n8n for a non-technical operator: Make, and 2.35 is the first release where that gap is close enough to name.

What changed since July, and what did not

The July review covered 2.26. Between then and 2.35.5 the buyer-relevant changes cluster on the agent builder, not the core engine. The table below is the delta that matters to a purchase decision.

Capability	July (2.26)	August (stable 2.35.x)
Agent test runs	None; you deployed and watched	Run a test execution of the agent before deploy; broken nodes flag in the test loop (2.35, Aug 11)
Human-in-the-loop	Build a manual wait node by hand	Per-tool review step: gate all tools or selected ones, approve via Slack / Telegram / n8n Chat (n8n HITL docs)
Cost visibility	Read the provider bill after the fact	Local token counting shows an estimated per-run budget in the editor (2.35 release notes)
Agent chat channels	Slack, Telegram	Slack, Telegram, and now Discord (2.35, Aug 11)
Cloud pricing	Starter 20 EUR, Pro 50 EUR	Unchanged (n8n pricing)
Self-hosting model	Docker, your ops burden	Same

Two of those rows are the whole story. Test runs mean you can execute an agent workflow against generated data and catch a misconfigured node before it answers a real user. Human review moved from a pattern you assembled to a toggle you set. In July, gating an agent's risky action meant wiring a wait node into the flow and handling the resume yourself. Now n8n's docs describe a review step you connect tools to: the agent pauses when it wants to call a gated tool, sends an approval request showing which tool and what parameters, and either executes on approval or cancels and tells the model it was denied (n8n HITL docs).

One tier down, the pre-release track adds more, but do not buy on it yet. n8n 2.36 (currently 2.36.3, pre-release) brings LangSmith debug export for agent traces, Slack as a first-class agent action, an MCP auto-expose toggle, and per-principal agent sandbox isolation for multi-tenant Cloud. Those are strong. They are also unshipped on the stable channel as of August 20, so this review scores 2.35 and flags 2.36 as a reason to check back.

Pricing: current euros, approximate dollars

Cloud pricing did not move since July. n8n bills in euros, annually, and the annual commitment saves 17% over monthly (n8n pricing, fetched 2026-08-20). The USD column below converts at roughly 1 EUR = 1.08 USD, an approximate mid-to-late-2026 rate, so read the dollar figures as ballpark for budgeting, not billing.

Plan	Price (EUR, annual, n8n pricing)	USD approx	Executions/mo	AI credits/mo	Hosting
Community Edition	0	$0	Unlimited	n/a	You self-host
Starter	20 EUR/mo	~$22/mo	2,500	2,300	n8n cloud
Pro	50 EUR/mo	~$54/mo	10,000	up to 13,700	n8n cloud
Business	667 EUR/mo	~$720/mo	40,000	included	n8n or self
Enterprise	Custom	Custom	Custom	included	n8n or self

Prices, execution caps, and AI credit allowances per n8n's pricing page, as of 2026-08-20.

Executions are the wall, and they arrive earlier than the plan names imply. The 2,500 on Starter reads as generous until you do the arithmetic on a single always-on automation. One workflow polling an API every fifteen minutes runs 4 times an hour, 96 times a day, about 2,880 times a month. That one poller alone clears the Starter cap before any of your other automations have fired. Pro's 10,000 absorbs most small-team loads; past that you are on Business at 667 EUR or you are self-hosting. This math has not changed since July, and it is still the reason the self-hosted path wins for anyone technical.

A managed VPS is the self-hosted escape. The Community Edition is a Docker image with no license fee and no execution meter, and a Cloudways droplet in the 11 EUR/mo range handles patching and backups so you are not also running a server by hand. That removes the execution cap entirely and keeps your data on your box.

Agent Builder: production-ready yet?

Closer than in July, but not all the way. The honest read follows.

Test runs lower the debugging floor. Before 2.35, an agent workflow's first real test was a live message, and a broken tool node failed in front of whoever sent it. The test run executes the agent against generated input so you see the failure in the editor instead. That is the single biggest reliability gain of the release, and it is the reason a cautious team can now prototype an agent without pointing it at real traffic on day one.

Per-tool human review is the second gain, and it changes what you can safely automate. Say your agent can read a CRM, draft a reply, and send an email. You want the read and the draft to run unattended but a human to sign off before anything leaves the building. In 2.35 you enable review on the send-email tool only. The agent runs freely until it reaches that tool, then it pauses and drops an approval request into Slack showing the exact recipient and body; a reviewer approves or denies, and on denial the model is told no and continues (n8n HITL docs). The gate is per-tool, so you are not forced to babysit the whole run to guard one irreversible action. n8n's own guidance is to start with review on and dial it back as confidence grows.

Token counting closes the third gap. The editor now estimates the token budget for an agent run before you execute it, so a run that would burn a fortune in context is visible before it bills, not after.

What is still missing is the reliability layer, same as July. The AI Agent node hands you the loop; it does not absorb LLM failure for you. Malformed model output can still halt a run, and a dependable production agent still means building the retry path and the fallback branch yourself. And the framing matters: n8n's docs label the whole "Build and manage agents" area Preview, and the 2.35 features are eight days old as of this writing. Turn them on in staging. Do not bet a customer-facing SLA on a v1 the week it ships.

MCP integration depth

n8n's MCP server is still the widest competitive moat, and it is where 2.36 will matter most once it lands. In July, n8n's native MCP server already let Claude, ChatGPT, or Cursor build and run a workflow from a plain-text prompt. The pre-release 2.36 track adds a settings toggle that auto-registers new workflows to the MCP server, so you skip the manual registration step, plus an MCP tool that reads structured rows from a workflow's data tables and a version-diff call (n8n releases). No other automation platform auto-exposes new workflows to an MCP server with a checkbox.

If you are wiring n8n into an agent stack that speaks MCP, the 2026-07-28 stateless MCP spec is the context worth reading first, because it changes how those server connections hold state. The auto-expose toggle is the feature to watch. It is also the clearest signal of where n8n is steering: from an automation tool with an agent node toward an agent platform with automation underneath. Just remember the toggle is pre-release today. The MCP server itself remains a public preview, not GA.

The verdict, per persona

The rating stays 4.5, and the reasoning is one sentence: the agent tooling got materially better, but every headline addition is a v1, a Preview, or a pre-release, so the score that already reflected n8n's self-hosting economics holds rather than climbs. The pick, by buyer:

Solo dev, comfortable with Docker. Run the free Community Edition on a managed VPS. An 11 EUR/mo Cloudways box gives you unlimited executions and the full Agent Builder for zero license cost, and your existing Claude or Cursor subscription becomes the workflow builder through the MCP server. Same call as July, and 2.35 only sweetens it.

Ops team, 2,500-plus executions a month, no one to own uptime. Buy Pro at 50 EUR/mo (pricing). You are paying to make patching and backups someone else's job, and the unlimited-active-workflows change keeps a portfolio of automations viable on one plan. If you have even one engineer who can run a container, self-host instead and put the difference into monitoring.

Non-technical operator who will not touch a container. Make still wins, but this is the first release where the gap is close. The 2.35 test run and the one-toggle review gate are exactly the friction reducers that were missing, and they point straight at this buyer. It is not enough yet: the builder still assumes you can reason about nodes and credentials, and the agent area is Preview. Give it two or three more versions before this flips.

Enterprise with multi-tenant agents. The credential-isolation gap that kept n8n Cloud out of multi-tenant deployments gets a fix, per-principal agent sandbox isolation, but that specific item is in the 2.36 pre-release, not stable. If this is your blocker, the answer is "almost, wait for 2.36 to go GA," not "yes today." Until then, Business or Enterprise on self-hosted infrastructure keeps tenant data on hardware you control.

Agent-first team that wants an MCP-native workflow. n8n is the clearest choice in the category. Nothing else pairs a first-party MCP server, a self-hostable agent builder, and the pending auto-expose toggle. This is the buyer n8n is now built for.

What to watch

Two triggers would move this review. First is 2.36 reaching stable, which turns LangSmith trace export, the MCP auto-expose toggle, and per-principal sandbox isolation from "check back later" into shipped features, and would make the enterprise call a yes. Second is the MCP server graduating from public preview to GA, which is what a customer-facing team needs before it trusts generated workflows past the draft stage. Until then, 2.35.5 is the version to run, and the pick stands.

Ready to start? Stand up the self-hosted Docker image from n8n if you want the pricing argument to hold, or run it managed on Cloudways to get the unlimited-execution economics without owning the ops stack.