No-Code AI Agent Builders Compared (2026): 7 Tools, One Shutting Down

I get asked which no-code AI agent builder to use about once a week, usually by someone who has already picked one and wants permission. My honest answer in late 2026 is that the canvases have converged and the bills have not. Every tool on this list can connect a model to a few tools and a memory store. What differs is the unit they charge you in, whether you can run it on your own machines, and whether it will still exist next year: OpenAI's Agent Builder, launched to a lot of noise in October 2025, shuts down on 30 November 2026. This post compares seven no-code AI agent builders on the things that decide the bill and the risk, works one message through each pricing model, and ends with where I would write code instead.

What counts as a no-code AI agent builder in 2026?

An agent, for this post, is a loop: a model reads the request, decides whether to call a tool, reads the result, and repeats until it can answer. A no-code builder gives you that loop as a box on a canvas, plus a catalogue of connectors you can hand it as tools, plus somewhere to keep conversation memory. The n8n AI Agent node is the clearest example: you connect a chat model, one or more tool sub-nodes, and optionally memory, and the agent chooses which tools to call. Zapier, Make, Gumloop and Relevance AI sell the same shape with different packaging. Copilot Studio wraps it in Microsoft 365 identity and licensing.

What none of them change is the arithmetic of agents. A workflow runs the steps you drew. An agent runs as many steps as the model decides to, which is why the billing unit matters more here than it did for classic automation. If you want the background on loops versus workflows, I wrote it up in the frameworks comparison.

The seven builders compared: price, billing unit, self-hosting

These are the numbers on each vendor's own pricing or documentation page as of 31 October 2026. Where a vendor does not publish a number, the table says so rather than guessing.

Builder	What you are billed for	Entry paid tier	Self-host	Notable limit	Source
n8n	Workflow executions (one per run, regardless of steps)	Starter EUR 20/mo annual, 2,500 executions	Yes, Sustainable Use License	5 concurrent executions on Starter	n8n pricing
Zapier Agents	Activities (each tool call, lookup or search)	Pro from about $33.33/mo annual, 1,500 activities	No	40 activities per run on paid plans	Zapier pricing
Make AI Agents	Credits: 1 per operation, plus tokens on Make's AI Provider	Plan prices not shown when I checked	No	Own API key only on paid plans	Make credits
Gumloop	Credits at $0.005: list-price tokens and tools, plus 8% fee	Pro $37/mo, 20,000 credits	VPC on Enterprise	Agents stop at the allowance unless pay-as-you-go is on	Gumloop pricing
Relevance AI	Actions (one per tool run, even failed) plus vendor credits	Public page shows Enterprise only; free plan retired	No	Action top-ups $80 per 1,000	Relevance docs
Copilot Studio	Copilot Credits: 5 per agent action, 2 per generative answer	$200/mo per 25,000-credit pack	No	Agents blocked at 125% of prepaid capacity	Microsoft billing rates
OpenAI Agent Builder	API tokens, no separate SKU	n/a	No	Shuts down 30 Nov 2026	OpenAI deprecations

Two rows deserve a comment. Relevance AI's pricing page now shows a single Enterprise card with "Talk to sales", and its docs say the Free plan "has since been retired for new signups and existing accounts". Pro and Team still exist in the docs, but I could not find their current prices on a Relevance page, so I have left them out. Make's pricing page did not render plan prices when I fetched it, so the only Make figures I am confident of are the credit rules in its help centre.

How each builder bills an agent, worked through one message

Take a modest internal helpdesk agent: 1,000 messages a month, each answered with three tool calls and one generated reply, on a non-reasoning model. Here is what each pricing model does with that, read straight from the published rules.

On n8n Cloud the Chat Trigger docs are explicit that every message runs the workflow and uses one execution. Tool calls inside the run do not add executions, so 1,000 messages is 1,000 of Starter's 2,500. You bring your own model credentials, so the token bill goes to OpenAI, Anthropic or whoever separately.

On Zapier, an activity is any billable step an agent takes: using a trigger, answering from a knowledge source, running an action, searching the web. Three tool calls is at least three activities, so the month is 3,000 or more, double the Pro allowance of 1,500. Zapier's own troubleshooting article lists the ways this grows: triggers that return ten items, knowledge lookups, and agent-to-Zap loops where the agent edits the data that triggers it.

On Copilot Studio, the billing table charges 5 credits per agent action and 2 per generative answer, so each message costs 17 credits and the month costs 17,000, inside one $200 pack. Switch to a reasoning model and Microsoft adds 10 credits per 1,000 billed tokens on top; its own worked example prices a single follow-up question at 141 credits for the reasoning tokens alone. Employee-facing use by people who already hold a Microsoft 365 Copilot licence is billed at no charge, which is the real reason Copilot Studio wins inside Microsoft shops.

On Gumloop, credits cover list-price tokens, compute and tool calls (minimum one credit per call), plus an 8% orchestration fee. Three calls is 3.24 credits before tokens, so 3,240 a month plus the token cost, well inside 20,000. Bring your own model key and the token credits go to zero but the fee doubles to 16%, charged on what the task would have cost without your key.

On Relevance AI, each tool run is one Action even if it fails, so the month is 3,000 Actions. On Make, each operation is a credit and Make's AI Provider converts tokens to credits at model-dependent rates (18,080 input tokens per credit on GPT-5 nano, 180 on Claude Opus 5); with your own key you pay Make for operations only.

The lesson from the grid is not that one vendor is cheap. It is that per-step billing punishes exactly the behaviour that makes agents useful. An agent that checks its work with one extra tool call costs a third more on Zapier or Relevance, 29% more on Copilot Studio, and nothing extra on n8n, where you pay per run and the model provider bills tokens. Choose the unit that matches how chatty your agent will be, then put a hard cap on iterations whichever tool you pick.

Can you self-host a no-code AI agent builder?

Of the seven, only n8n gives you the whole product on your own hardware. Gumloop lists a virtual private cloud as an Enterprise option; the rest are SaaS only, and Copilot Studio's pricing page says plainly that an Azure subscription is required. Note what n8n's licence actually says: the Sustainable Use License permits use "only for your own internal business purposes or for non-commercial or personal use". Running agents for your own company is fine. Reselling n8n-powered agents as a hosted product is not, without a commercial agreement.

If you need a licence the OSI recognises, the open-source alternatives are mostly LLM-app builders rather than integration platforms. Langflow is MIT and shipped v1.13.0 on 22 October. Dify shipped 1.18.0 on 23 October under a modified Apache 2.0 licence that forbids running a multi-tenant service without permission. Flowise, which a lot of 2024 tutorials still recommend, had its repository archived in August 2026 after a final 3.1.4 release.

Self-hosting moves the security work to you, and with agent builders that work is real. They hold every credential the agents can use, and they execute code. This is the check I run before I trust any n8n install:

# Count n8n's GitHub security advisories published in 2026, by severity.
# Run on 31 Oct 2026 with gh 2.49.2; output shown below the command.
gh api --paginate "repos/n8n-io/n8n/security-advisories?per_page=100" \
  --jq '.[] | select(.published_at >= "2026-01-01") | .severity' | sort | uniq -c
#   19 critical
#   97 high
#    2 low
#  105 medium

# The current stable release, which is what you should pin (never :latest)
gh api repos/n8n-io/n8n/releases/latest --jq '.tag_name + " " + .published_at'
# n8n@2.45.7 2026-10-30T08:24:56Z

That is 223 advisories in ten months, 19 of them critical. Most are fixed quickly, and a busy project publishes more advisories than a quiet one, but the number tells you the patch cadence you are signing up for. I covered the exploited n8n and Langflow CVEs, and the exposed-host counts, in the workflow automation security post. The short version: pin a version, patch weekly, never expose the editor to the internet, and run it with the same discipline as any other system that holds production keys.

OpenAI Agent Builder is shutting down: what that teaches about platform risk

OpenAI announced AgentKit on 6 October 2025, with Agent Builder in beta and, per VentureBeat's launch coverage, pricing "included in the standard API model pricing". On 3 June 2026 OpenAI's deprecations page announced the shutdown for 30 November 2026, pointing users to the Agents SDK or ChatGPT Workspace Agents. The Evals platform, announced alongside it, goes read-only today and shuts down the same day. ChatKit survives.

Agent Builder never left beta, so nobody should be shocked. But it is the cleanest example of the risk every no-code agent builder carries: the workflow lives in the vendor's format, and when the product goes, you rebuild. Agent Builder had an escape hatch, exporting workflow code for the Agents SDK, which is more than most canvases give you. If you are on it, export now rather than on 29 November.

The same question applies to the survivors. n8n workflows are JSON you can keep in git and import into any n8n instance you run. Zapier and Make let you look at the logic but not run it anywhere else. Copilot Studio agents live in Power Platform solutions, portable between Microsoft environments and nowhere else. Before you build something people depend on, decide what you would do if the builder disappeared. It is a ten-minute conversation that saves a quarter.

When should you write code instead of using a no-code agent builder?

I use no-code builders for the agents that are mostly integration: read a ticket, look something up in two SaaS tools, post a summary. They are faster to build than code and the connector catalogue is the real product. I write code when any of these is true:

• The agent is the product. If customers talk to it, you need tests, evals in CI, versioned prompts and a deploy pipeline. Canvases are bad at diffs and worse at code review.
• Per-step billing gets in the way of quality. When you find yourself removing a verification step to save activities, the tool is choosing your architecture.
• You need custom control flow. Human approval mid-run, parallel sub-agents, checkpoint and resume. LangGraph or the OpenAI Agents SDK do this natively; I walked through one version in the human-in-the-loop guide.
• The agent runs code or touches untrusted input at scale. Then isolation, egress control and audit logs matter more than the canvas. I run PandaStack, a Firecracker microVM cloud for agent code execution, so weigh my bias accordingly, but whatever you use, model-generated code should not run in the same process that holds your API keys.

The middle path is underrated: build the agent in n8n, expose the hard parts as tools you write in code (an HTTP endpoint or an MCP server), and let the canvas do the plumbing. The step-by-step n8n agent guide shows exactly that, with Postgres memory and an MCP tool.

FAQ

What is the best no-code AI agent builder in 2026?

There is no single best one; it depends on where you already live. Inside Microsoft 365, Copilot Studio wins because licensed employees use agents at no extra charge. If you need self-hosting or chatty agents with many tool calls, n8n's per-execution billing is the most predictable. For the widest SaaS connector catalogue with the least setup, Zapier.

Is n8n free for building AI agents?

The self-hosted Community edition costs nothing in licence fees under the Sustainable Use License, which allows internal business use but not reselling it as a service. You still pay your model provider for tokens and you run the server. n8n Cloud starts at EUR 20 a month billed annually for 2,500 executions.

Is OpenAI Agent Builder being discontinued?

Yes. OpenAI's deprecations page lists Agent Builder for shutdown on 30 November 2026, announced on 3 June 2026. The recommended paths are the Agents SDK or ChatGPT Workspace Agents, and ChatKit is not affected.

How much does a Zapier AI agent cost?

Zapier Agents are billed in activities, separately from Zap tasks: 400 a month free, 1,500 on the Pro tier from about $33.33 a month billed annually. Each tool call, knowledge lookup or web search is an activity, and a single run is capped at 40 on paid plans, so estimate tool calls per run before you commit.

This quarter, if I were picking for a small team, I would put the integration-heavy agents on a pinned, self-hosted n8n behind a VPN, give each agent a maximum iteration count and its own narrowly scoped credentials, and write the one customer-facing agent in code with evals from day one. Anyone on Agent Builder has a month. Export the code this week.

Related: Workflow Automation in 2026: n8n Published 196 Advisories in Nine Months, 59,558 Hosts Sat Exposed, and the Code Alternative Raised at $12.55B, AI Agent Frameworks Compared (2026): LangGraph, CrewAI, ADK and 7 More and No-Code Didn't Die. It Sold for 2.7x ARR While Its Replacement Raised at 27x.