Let agents work for hours — inside a workflow you control.

neonloops is a visual builder for agentic workflows — an AI workflow builder where the AI does more than answer. You describe the job in plain language and the workflow gets assembled for you, or you build it yourself on the canvas: what triggers it, what order the steps run in, where it stops for your approval. Inside it, AI agents do the part that can't be scripted — read the 400 invoices, work out which prices moved — and the workflow does the same right thing with whatever they produce, every time.

An agentic workflow is an automation where AI agents decide how to get the job done, inside a structure you defined. A traditional workflow follows fixed if-this-then-that rules, and a single AI step just answers a prompt. An agentic workflow gives an agent an open-ended job — go through these contracts and flag the risky ones — lets it work its own way through hundreds of steps, and then handles the result deterministically: file it, assign it, or hold it for a human.

Because one AI step can only answer once. It reads what you hand it and returns a paragraph — it can't go looking, change its approach when the first answer is wrong, or keep working for an afternoon. Real jobs need that: reading a year of invoices, checking each one against its contract, chasing the ones that don't match. An agent works through that unpredictably, which is exactly why it needs a workflow around it.

Because you'd have to ask — every time, and only when you remember to. An assistant answers when you open it and type. A workflow starts on its own: on the 1st of the month, or the moment an account cancels. It runs the same steps in the same order, pauses where you told it to wait for you, and leaves a record of what it did.

Getting a good answer isn't the hard part any more. Not having to be the one who asks for it is.

Runs are durable today: they survive restarts, wait as long as they need to for an approval, and pick up where they left off — so a single run can legitimately span days. Agents working autonomously for hours inside those runs is what we're building now, and it's the point of the whole product. We'd rather tell you which half is which than let you find out later.

For the model steps — sorting, extracting, drafting — you can use OpenAI, Anthropic and Google, and mix them inside the same workflow: a cheap model to triage, a stronger one to reason. Model access is handled for you: we run the provider accounts behind every step, so there is no API key to go and find, paste in, or rotate later.

Agents are the exception, on purpose. We're building the long-running agent work on Claude's managed agent infrastructure, so agent steps use Claude models — that's the trade for an agent that keeps working for hours instead of answering once. Either way, every run reports the tokens it used, per step, so a long job is never a black box.

For the time your work is actually working — the minutes a workflow or an agent spends genuinely doing something for you. Not per seat, not per run, and there are no tiers to pick between: every feature is included for everyone.

Three things fall out of that, and they are the reassuring part. Waiting doesn't bill: a workflow parked three days on a human approval bills nothing for those three days. Runs are counted but never charged, so a trigger that fires 400 times and exits every time costs nothing — a run one of those events actually starts bills like any other. And you set the ceiling yourself: when your work reaches it, we stop it, and usage is never charged past it — VAT, where it applies, comes on top. Everything you have already built stays open to read. Tables, files and skills export any time; workflows and run history export on request.

There's no trial: a workspace adds a card and is priced per billable minute from the start, against the ceiling you set. The rate is on your billing screen before you ever start — nothing about what you're charged is a surprise.

How pricing works

Work that repeats, needs real judgement, and eats hours, days or weeks depending on the month. Auditing a year of supplier invoices for price rises. Reading every contract for the ones that auto-renew. Reviewing three thousand product listings. Reconciling a month of payments and chasing the mismatches. Writing up why an account churned.

The test is whether you could write the steps down in advance. If you can, an ordinary workflow will do it and you don't need an agent. Agents earn their place where the path changes with whatever they find — every invoice is laid out differently, every contract words the clause differently — and where the job runs long enough that nobody wants to sit through it. That unspecifiable middle is exactly the part they take over.

No. You describe what you want in plain language and the workflow gets built for you; you can also drag the steps yourself on the canvas, and say what each one should do in plain language too. Every node has a developer view when you want exact control over a prompt or a payload, but you can design, test and run a workflow without writing any code.

Yes, and this is the part that makes agents usable. You can put an approval step anywhere: the run pauses, waits for you — for minutes or for days — and only continues once you say so. Nothing goes out, gets paid, or gets published without the sign-off you asked for, no matter what the agent decided.

Yes. neonloops has an MCP server, so you can keep working where you already are — connect Claude, Cursor, ChatGPT, or any other MCP client, and it can build or run a workflow, check on a run, approve a step, or read a table for you.

See how to connect