OpenAI Agent Builder Shuts Down on 30 November: 7 No-Code Alternatives, Compared

OpenAI is switching Agent Builder off. The banner on its own documentation says the product "is scheduled to shut down on November 30, 2026", and that existing users "can continue using it during the transition window". The deprecations page dates the notice to 3 June 2026.

Agent Builder shipped at DevDay on 6 October 2025, inside AgentKit, which Sam Altman introduced as "a complete set of building blocks available in the open AI platform designed to help you take agents from prototype to production". TechCrunch's write-up that day relayed the in-house pitch: like Canva, for building agents. OpenAI described the thing itself as "a visual canvas for building multi-step agent workflows" where you "drag and drop nodes for each step in your workflow". OpenAI Agent Builder explained covers what the product was and what it cost, if you want the background.

Fourteen months, launch to shutdown. The death notice arrived when it was eight months old.

If you write code, this is an afternoon of annoyance. OpenAI exports your workflow as Agents SDK code, you open an editor, you carry on. If you do not write code, that export is not a migration. It is a wall, written in TypeScript.

The wall has a second course of bricks on top. OpenAI's own no-code destination, Workspace Agents, needs "a ChatGPT Business, Enterprise, or Edu workspace with access to workspace agents and permission to create agents". An individual account does not qualify. Neither does Plus. So the visual canvas has no visual successor you can reach without a company buying you a licence.

Seven places to go. Every price came off the vendor's own page on 12 September 2026, in the currency they publish it in. Mine is last, because I sell it.

What OpenAI is actually shutting down, and when

Three things died in one announcement, and most coverage has only counted one of them. Check whether you were quietly using the other two before you plan anything.

• 3 June 2026. The deprecations page records it plainly: "On June 3, 2026, we notified developers using Agent Builder that the product is being deprecated."
• 31 October 2026. Existing evals become read-only. You can look at them. You cannot add to them.
• 30 November 2026. Agent Builder shuts down. The Evals dashboard and API shut down on the same day. So does the v1/prompts API behind reusable prompts.

Reusable prompts are the quiet casualty. That endpoint gets called from production code by people who never opened the canvas, and it stops answering on 30 November alongside everything else. If your engineer set that up eighteen months ago and left, nobody in your company currently knows it is there.

Then there is the part that reads like good news.

ChatKit survives. OpenAI's wording is two words long: "ChatKit remains available". That is the most dangerous sentence in the whole announcement, because a ChatKit widget embedded on your marketing site is a front end, and the Agent Builder workflow behind it is the back end. On 30 November the front end keeps rendering perfectly. The thing it talks to is gone.

A chat box that has vanished gets reported by a customer within an hour. A chat box that loads, accepts a question and then does nothing useful can sit there for a month collecting abandoned conversations. I could not find another write-up of this shutdown that costs out the orphaned widget.

How I compared them

Five questions, asked of every candidate. They sort this field faster than a feature grid does, because in a forced migration the features are rarely what hurts you.

• What does the meter count? This is the real decision variable and almost nobody writes it down. The same workload is a workflow execution in one place, a module action in another, and two simultaneous meters in a third.
• What is the starting price, from the vendor's own page? Read on 12 September 2026, in the currency published, with no conversion applied.
• Is there a free tier you can actually rebuild inside? A one-off allocation of 200 credits and a monthly allowance of 1,000 are different products wearing the same word.
• Is there an Agent Builder importer? Asked of eleven vendors, including the four I ended up rejecting.
• Is this still a visual canvas, or has it quietly become something else? You are here because you wanted to see the boxes and the arrows.

Here is exactly what I did, and it is worth being precise about, because a comparison post that implies a lab is lying to you. On 12 September 2026 I read every vendor's own pricing page and the relevant parts of their documentation. I read OpenAI's Agent Builder guide, its deprecations page and its migration guide the same day. I then searched all eleven vendors' docs and blogs for any Agent Builder importer, converter or migration tool.

I ran no tests. I did not deploy anything, migrate anything or benchmark anything, and where a page would not load I have said so in the section it belongs to and left the figure out.

The importer search came back empty. None found as of 12 September 2026, at any of the eleven.

Quick comparison

Every figure read on the vendor's own page on 12 September 2026, except Gumloop's, which came from Gumloop's documentation because its pricing page would not render. n8n publishes in euros and no conversion is applied. Dify and n8n Starter quote annual billing. Gravity's row is my own product. Nothing in this table was tested; it records what the vendors publish.

1. Dify: the closest structural match to Agent Builder

If you want the same shape of tool, start here. Dify is a visual, LLM-native orchestration canvas: nodes, branches, model calls, retrieval, the lot, arranged on a board you can see. Agent Builder used the same mental model, which matters more than any single feature when you are rebuilding from memory.

Professional is $590 per workspace per year, billed annually, which Dify's page frames as saving 17% against monthly. Call it about $49 a month. The plan carries 5,000 message credits a month, and they belong to the workspace, so a two-person team and a one-person team draw from the same pool. Team is $1,590 per workspace per year for 10,000 credits a month, 50 members and 200 apps.

Two things to know before you commit. Professional is annual-billed only, so the smallest real commitment is $590 up front. And it caps at three members, which is a low ceiling if the reason you used a visual canvas was that several non-technical people could read it.

The free Sandbox tier needs reading carefully. You get 200 message credits once. They do not refresh. One workspace, one member, five apps. You can evaluate the product on it. You cannot run anything on it.

Dify's DSL import reads Dify files and nothing else

Dify does have import and export, in a YAML DSL format, and on a page about migrations that looks like hope. It is not. Dify's own app-management documentation scopes the DSL to moving apps between Dify instances and Dify workspaces. Your Agent Builder export is Agents SDK code in TypeScript or Python, and Dify does not read that.

Worth having anyway, for the next time. An exportable definition is the difference between a rebuild and a copy when the following platform decides it has other plans.

Best for

An Agent Builder workflow you want to rebuild as nearly the same diagram, run by up to three people.

Pricing

Sandbox free with 200 one-off message credits. Professional $590 per workspace a year, about $49 a month, 5,000 credits a month, 3 members. Team $1,590 per workspace a year, 10,000 credits a month, 50 members, 200 apps. Community Edition free to self-host. As of September 2026, from dify.ai/pricing.

Free tier

Yes, with 200 one-off message credits. Enough to evaluate on, and that is all.

Verdict

The shortest conceptual distance from Agent Builder, at the cost of an annual commitment and a three-seat cap. If you want the picture back, this is the picture.

2. n8n: the fairest meter in the category, if someone technical is nearby

n8n bills one execution per workflow run, and it does not care how many nodes are inside that run. A three-node workflow and a forty-node workflow cost exactly the same. For anyone rebuilding a multi-step Agent Builder graph, that is the single friendliest pricing model on this page.

Starter is 20 euros a month billed annually, for 2,500 executions with unlimited steps. Pro is 50 euros for 10,000. Business is 667 euros. Note the prices are in euros. n8n does not convert them for you.

A 180-second ceiling decides whether the trial is usable

There is no free n8n cloud plan. There is a trial, and its limits are specific: 1,000 executions, 5 concurrent executions, a 180-second execution timeout, and 1 team project.

Three minutes is the number to stare at. An agent loop that calls a model, reads a document, calls the model again and writes somewhere will regularly run past three minutes, and on the trial it does not fail gracefully at the end. It gets cut off. So you can evaluate simple automations on the trial and you cannot properly evaluate the long agentic workflows that Agent Builder existed to build, which is an awkward place for a trial to stop.

The honest answer to that is the Community Edition, free and self-hosted, with no timeout that anyone else sets. The honest problem with that answer is that you are a no-code buyer and somebody now owns a server.

With one person in the building who enjoys this kind of work, n8n gets you the most capability per euro of anything here. Without them, 20 euros a month buys a canvas that is harder to read than the one you lost and nobody to maintain it.

3. Make: the cheapest paid canvas, and the meter that punishes long workflows

Nine dollars a month, 5,000 credits. No paid number on this page is smaller, and Make's visual builder is genuinely pleasant: routers, filters and iterators drawn as shapes, branching you can follow without reading anything.

Then read what a credit is. Make says it outright: "Each module action in your scenario, like adding a Google Sheet row or fetching Gmail account data, counts as one credit."

The same workflow, metered two ways

Take one workflow and price it on both models, using nothing but the published rates. Twelve steps, running 500 times a month, which is a fairly ordinary rebuild of a fairly ordinary Agent Builder graph.

On n8n that is 500 executions. Starter covers 2,500, so you are at a fifth of your allowance, paying 20 euros a month, and you could triple the length of the workflow without the bill moving at all.

On Make that is 12 module actions multiplied by 500 runs, which is 6,000 credits. The $9 plan carries 5,000. You are over the line in month one, before anyone has added a step.

Same work. One meter barely notices it; the other needs a bigger plan. This is why "which is cheapest" is an unanswerable question until you know how long your workflow is, and it is the single most useful thing to carry away from this comparison.

Make's free plan gives you 1,000 credits a month, with a catch worth knowing before you plan an evaluation: it enforces a 15-minute minimum interval between runs. Anything that needs to react quickly cannot be tested properly on it. There is no self-hosting option at all.

Best for

A short, high-frequency scenario where the bill was always going to be the sticking point.

Pricing

Free with 1,000 credits a month and a 15-minute minimum interval between runs. Paid from $9 a month for 5,000 credits, sized upward from there. As of September 2026, from make.com/en/pricing.

Free tier

Yes, 1,000 credits every month.

Verdict

Cheapest way back onto a canvas, provided your workflow is short. Count your steps and multiply before you believe the $9.

4. Gumloop: AI-native nodes, one paid tier, and a pricing page that would not load

Gumloop is built the way Agent Builder was built, with the model at the centre. Any node can be an AI step, and the canvas assumes that is the normal case. If what you liked about Agent Builder was chaining model calls with a bit of logic between them, this is the closest match after Dify.

Pro is $37 a month and carries 20,000 credits. Gumloop prices a credit precisely: "1 credit = $0.005", and "$1 buys 200 credits". Overage runs at the same $0.005 a credit. Credits do not roll over from month to month unless you are on Enterprise, so an underused month is money gone.

Why these numbers come from Gumloop's docs and not its pricing page

gumloop.com/pricing is JavaScript-rendered and returned nothing usable on three separate attempts on 12 September 2026. So these figures come from Gumloop's documentation at docs.gumloop.com, which states the credit price and the Pro allowance directly. Same company, different surface, and you should know which one.

The free tier has the same problem and a worse ending. Gumloop's docs describe a 14-day free trial of Pro, and nothing else. Several third-party sites claim a permanent free plan exists, and they give three different numbers for it. So I am not publishing one. Assume a trial, and if you find a free plan when you sign up, treat it as a bonus.

Start the trial the week you start rebuilding. Fourteen days disappears quickly when you are reverse-engineering your own workflow from a code export, and best Gumloop alternatives covers the wider field if the single paid tier turns out to be the wrong size.

5. Zapier Agents: the biggest catalogue, and two meters running at once

Zapier is here for one reason, and it is a good one. If your Agent Builder workflow reached into some app that nothing else on this page has heard of, Zapier probably has it on the shelf. Nobody matches the connector catalogue.

It is also the furthest from what you had. Zapier Agents is configured as a step list plus natural-language instructions. There is no free-form node canvas to draw on. Closest on connectors, furthest on interface metaphor, and you should know that before you invest a weekend.

Two meters run during one agent run

The core Zapier plan meters tasks. Professional is $19.99 a month billed annually for 750 tasks a month, or $29.99 billed monthly. A task is "counted whenever Zapier successfully completes a unit of work for you".

Agents is a separate add-on with its own meter entirely. Free gives you 400 activities a month; Pro gives you 1,500 at about $33.33 a month billed annually. Zapier defines an activity as "any billable action an agent takes: using a trigger, answering from a knowledge source, running an action, web browsing or scraping, a web search, or a message sent via the Chrome extension".

And Zapier is explicit about how the two relate: "Add-on billing is separate from your core plan." An agent that does real work can therefore draw down agent activities and core tasks in the same run. Budget both, because a single number will mislead you.

Nobody has built an importer, including the vendor best placed to

Zapier has the largest connector-migration operation in this category and the most to gain from catching displaced Agent Builder users. It has not shipped an Agent Builder importer either. None of the eleven vendors I searched had one on 12 September 2026, which tells you something about how many no-code customers the product is thought to have had.

A missing connector that would sink the whole project is a good enough reason to choose Zapier, and the interface change is the price of that. Two meters, and the add-on bills separately from the core plan. Price both before you commit.

6. Microsoft Copilot Studio: sensible if Microsoft already has your money

Copilot Studio is the enterprise answer, and its pricing reflects that. A prepaid capacity pack is $200.00 a month for 25,000 Copilot Credits. Copilot Studio also comes bundled inside Microsoft 365 Copilot at "$30.00 user/month, paid yearly", which is how most companies will actually end up with it.

Credits pool at the tenant level, which is the right shape for a company and the wrong shape for one person. Microsoft is also unusually honest about how the meter behaves: "The number of Copilot Credits counted for each response or action depends on the complexity of the task the agent completes." Read that twice. The unit is not fixed, so the same run can cost different amounts depending on what it had to do.

Two more clauses from Microsoft's own billing documentation belong in your forecast. Unused credits "don't carry over to the next month". And exceeding your capacity has teeth: "technical enforcement applies and can result in service denial".

The trial licence will not let you publish

This is the line that decides it for most people on a deadline. Microsoft's documentation states: "The trial license gives you access to Copilot Studio to create agents. You can test your agents using the test chat panel. However, you can't publish the agent."

So you can rebuild your entire Agent Builder workflow inside the trial, confirm it works in the test panel, and still not have anything running. Getting to a live agent requires a paid licence first. If your plan for November was "evaluate on a trial, cut over on the 29th", that plan does not work here.

One caveat on my own numbers: several secondary sources quote a pay-as-you-go rate per credit, and I could not confirm it on any Microsoft page on 12 September 2026, so it is not in this post.

7. Gravity: no canvas at all, which is either the point or the dealbreaker

This is my product. Discount everything in this section accordingly, and it sits last for that reason.

Let me be straight about the fit, because it is narrow. Every other option on this list asks you to rebuild a canvas. We do not have a canvas. There is nothing here to drag, nothing to connect with arrows, and no board to admire when it is finished. If the building was the part of Agent Builder you liked, one of the six above is a better home for you and you should go and pick it.

If the canvas was only ever the means, this works differently. You describe the job in a prompt box, in ordinary English. What comes back is a finished agent from the catalogue, published and maintained by an independent builder. You connect the accounts it needs, answer its setup questions, and it runs on schedule. There is no graph to redraw, which in a migration month is the entire argument.

The first agent is free and asks for no card. After that, Autopilot is $20 a month worldwide with a lot of usage included, ₹1,999 a month in India, and India also has Minipilot at ₹399 a month with a smaller allowance. If you run out of usage you can buy more from your account. One number a month, no meter to watch, which after reading the previous six sections you may find appealing.

What you give up when there is no canvas

Custom logic stitched across three systems in a specific order still belongs on a canvas, and if that describes your workflow then nothing I say here changes it. Our catalogue is also short, because Gravity only went live in September 2026, so there are jobs nobody has built an agent for yet. And there is no self-hosting.

The thing you gain is who owns the maintenance. On every other route here, you rebuild it this month and you keep it alive for the next two years. Here the builder who published the agent owns the repairs, and their income depends on it still working in March.

Best for

An Agent Builder workflow that existed to finish one recurring job, where the diagram was a means and the output was the point.

Pricing

First agent free, no card. Then Autopilot at $20 a month worldwide, ₹1,999 a month in India, with a lot of usage included. India also has Minipilot at ₹399 a month with a smaller allowance. Buy more usage if you run out. As of September 2026, on the pricing page.

Free tier

Yes. One agent, free, and it does not expire in fourteen days.

Verdict

The only route on this page with no rebuild in it, which is worth more in a deprecation quarter than in a normal one. Wrong answer if you wanted the boxes and arrows back. What is an AI agent platform sets out the category difference properly.

What does not survive the export

OpenAI's migration guide gives you four steps, and they are worth quoting exactly, because they take about ninety seconds and they are the only part of this that is easy: "1. Open your workflow in Agent Builder. 2. Select Code in the top navigation. 3. Select Agents SDK in the code dialog. 4. Select TypeScript or Python, then copy the complete export."

Now the part in the same document that nobody quotes.

"This process does not convert your workflow graph or guarantee that every behavior transfers unchanged." That is OpenAI telling you the export is lossy, in its own words, on its own page. The graph is the thing you built. The graph does not come.

"Connected apps, authentication, publishing, and permission configuration require separate review in ChatGPT." So every integration, every credential and every permission decision is a manual job, done again, one at a time.

"Workflows with strong determinism at their core may not migrate faithfully to a workspace agent." Read that as a warning specifically aimed at you. Strong determinism is what a no-code buyer builds: if this, then exactly that, every time. The more reliably your workflow behaved, the worse it travels.

Add the licence gate on top. The destination that guide points at needs "a ChatGPT Business, Enterprise, or Edu workspace with access to workspace agents and permission to create agents", which an individual or Plus account does not have.

And do not forget the widget. If a ChatKit chat box on your site fronts one of these workflows, it keeps rendering on 1 December while the thing behind it is gone. Put it on the list now, because it will not announce itself.

So here is what the export actually is. Treat it as a specification document. It is the most complete written description of what you built that will ever exist, and its only real use is to sit open on one half of your screen while you rebuild by hand on the other half. Read the code even if you cannot write it: the node names, the order, the conditions and the prompts are all legible to a non-programmer, and that is the spec.

Rebuilding without an importer: a checklist

Gartner predicted on 25 June 2025 that over 40% of agentic AI projects will be canceled by end-2027. The figure sits on a January 2025 poll of 3,412 webinar attendees, and analyst Anushree Verma's framing was about cost, unclear value and inadequate risk controls. Gartner's own site blocks our fetcher, so I am relying on the figures as corroborated across multiple outlets covering the release.

Agent Builder is one line inside that forecast. Plan on the assumption that your next platform could be another one.

1. Export every workflow this week. Four clicks per workflow, per OpenAI's guide. The canvas is live during the transition window, which means today you can still read the original while comparing it against the export. On 1 December you cannot.
2. Screenshot the graph before you lose it. The export does not convert your workflow graph, in OpenAI's own words, so the picture only exists on your screen. Capture every canvas at full width, including the branch conditions. This takes ten minutes and saves a fortnight of trying to remember what fed what.
3. List your workflows and cut the list in half. Most people have three that matter and nine that were experiments. A forced migration is a decent excuse to let the nine go, and every one you keep is a rebuild you have signed up for.
4. Write down every connected app and credential separately. OpenAI says authentication and permission configuration require separate review, which means none of it travels. Note which account authorised what, because that knowledge lives in one person's head and they may be on leave in November.
5. Count your steps before you pick a platform. Twelve steps at 500 runs a month is 500 executions on n8n and 6,000 credits on Make. Same workflow, different plan tier. Do this arithmetic with your real numbers before you enter a card.
6. Check whether anything in production calls the prompts API. The v1/prompts endpoint dies on 30 November too. Grep your codebase, or ask whoever wrote the integration, because that failure has nothing to do with the canvas and will surprise you.
7. Find your ChatKit widgets. Search your site's source for the embed. Each one is a live chat box that will keep loading after its workflow stops, and it is the only failure on this list your customers will discover before you do.
8. Rebuild in parallel and compare real outputs. Agent Builder still runs during the transition window, so for a few weeks you can run both and diff the results. After 30 November your only reference is the invoice that got paid or the email that went out.
9. Revoke the OAuth grants from your side. A shut-down product does not necessarily release the access it holds on your Google, Slack or Notion accounts. Go into each of those apps' own security settings and remove it. Access grants outlive the products that requested them.

Steps one, two and seven are today's work and take under an hour between them.

Step eight is where your November goes, which is exactly why step three should be ruthless.

Which one should you choose?

Start with the shape of the tool. You are here because a visual canvas was taken away from you, and only four of these seven give you one back.

Dify is the closest match, full stop. Same metaphor, same LLM-native assumptions, and a YAML export that gives your next migration a starting file. The cost of admission is $590 a year up front and a three-member cap.

n8n is the right answer when one person in the building enjoys this work. One execution per run is the fairest meter in the category and it gets fairer as your workflow grows. Just do not judge it on the trial, where the 180-second timeout cuts off the exact workloads you need to test. Best n8n alternatives prices the same workload across the neighbouring tools if you decide the canvas itself was the problem.

Make wins on price only if your workflow is short. Nine dollars a month is real, and so is one credit per module action, and the second fact eats the first past about ten steps at 500 runs a month.

Gumloop suits the case where the model was doing the work and the wiring was just wiring. Budget for a 14-day trial, because a free plan is not what the documentation describes.

Zapier Agents is for the missing connector, and that is a perfectly good reason to choose it. Accept that you are trading a canvas for a step list, and that two meters means the real cost is closer to $53 a month than the $19.99 on the pricing page. Best Zapier alternatives using AI has the comparison from the other direction.

Copilot Studio makes sense when Microsoft already has your money and your identity provider. It makes no sense at all as a solo purchase at $200 a month for a capacity pack, and the trial's inability to publish makes a last-minute evaluation impossible.

My bias, stated plainly. If your workflow existed to finish a recurring job and the diagram was never the point, I would skip the rebuild entirely and pick Gravity, where the first agent is free and there is nothing to draw. If the diagram was the point, take Dify and enjoy it. Best no-code AI agent platforms covers the wider category if none of these seven fits.

Honourable mention: Lindy, priced per seat

Lindy is a good product in the wrong shape for this particular migration. Plus is $29.99 per user a month with 3,000 credits per user, and credits pool across the workspace. Lindy's page is clear that "Monthly credits refresh and do not roll over", and there is no permanent free plan listed on it as of September 2026.

Per-seat pricing is the issue. A solo operator rebuilding one workflow pays the same per head as an enterprise, and the credit pool grows only when you add people you do not need. It is worth a look if your Agent Builder workflow was really an inbox in disguise, and best Lindy alternatives has the detail.

Four I looked at and did not recommend

Rejections are more useful than recommendations in a forced migration, because they save you the afternoon you were about to spend.

Relevance AI. Its own page states that "The Free plan is retired and closed to new signups". A buyer who has just been evicted from a free visual tool cannot even try this one without paying, which disqualifies it on its first screen. The meter is also dual, Actions plus Vendor Credits, with overage at $80 per 1,000 Actions.

Flowise. The free tier is 100 predictions a month, which is enough for a demo and nothing more. Its genuine strength is self-hosting, and self-hosting is the specific thing a displaced no-code buyer is trying to avoid.

Langflow. It publishes no first-party managed price at all. Secondary reporting, which I could not confirm on a first-party page, says its DataStax-managed Astra hosting was discontinued in April 2026. Recommending a tool whose own managed host may have been shut down, to a reader fleeing a shut-down host, would be the worst call available in this post.

Google Vertex AI Agent Builder, now Gemini Enterprise Agent Platform. Our fetch of Google's pricing page failed on 12 September 2026, so I have no verified figures and no verified billing unit, and I am not going to guess at either. It was also renamed mid-2026, which makes its documentation a scavenger hunt. A no-code buyer on an eleven-week deadline should not be the one doing that archaeology.

One more name you will meet on other lists: Relay.app. Do not migrate to it. It is winding down, paid access ends on 14 September 2026, and any 2026 listicle still recommending it has not been updated since Relay's 16 July notice.

FAQ

Sources