Yes, you can build an AI agent for free, and a real one at that, one that takes over a recurring task. But the free paths have clear limits, and most articles on the topic don't mention them. Here's the honest overview: three paths, what each one really delivers, where it gets stuck, and when paying starts to make sense. All prices: as of August 2026.
For orientation first: what an AI agent even is and which use cases are worth it, you'll find on the AI agents page. This article only settles the money question.
Path 1: Claude Free
This is the path I recommend if you want to start without any tech setup. Claude's free plan includes Projects, meaning dedicated workspaces with fixed instructions and a knowledge base. That's exactly what you build an agent from: create a project, describe the job, upload your knowledge, test. The guide is in Build an AI agent with Claude.
What's free: a fully working agent for one task, with your context, no code, no credit card. For trying things out and occasional use, that's enough. You can test on real cases whether the agent actually takes work off your plate, and that's exactly what the free plan is for.
The limits: The message allowance is tight and resets in windows of a few hours. How many messages you get depends on the length of your conversations, and an agent with a well-filled knowledge base uses up more of it per message. On top of that, the number of projects on the free plan is limited; unlimited projects only come with Pro, and the strongest model stays reserved for the paid plans.
When paying makes sense: as soon as the agent works daily. Claude Pro costs 20 dollars a month, 17 dollars a month with annual billing, and above that there's Max starting at 100 dollars a month (as of August 2026). You don't need Max for your first agent; it only pays off once several agents or a whole team work for you daily.
Path 2: ChatGPT Free
With free ChatGPT you can also work in an agent-like way: put instructions into the chat, upload files, have it build results. For individual tasks here and there, that's perfectly fine, and if you use ChatGPT anyway, trying it costs nothing.
What's free: getting individual tasks done with the standard model, uploading files with daily limits, using existing GPTs from the catalog.
The limits: The feature that matters for agents is now missing entirely. Custom GPTs, meaning permanently configured assistants with their own instructions and knowledge, can no longer be newly created on personal accounts, and that goes for every personal plan, including Plus and Pro. Only Business, Enterprise, and Edu workspaces can still create new ones (as of August 2026). Already existing GPTs remain usable. Without your own GPTs, all you have on the free plan is the regular chat, and there you start from zero on every task: copy the context together, prompt, rework. Why exactly that is the bottleneck is covered in ChatGPT forgets everything.
For a permanent agent with its own knowledge, ChatGPT Free is currently the weakest of the three options. That's not a matter of taste, it's a matter of features.
Path 3: self-hosting n8n
n8n is a workflow platform: you build procedures as flowcharts, with an AI block at the points where text needs generating or a decision needs making. The Community Edition is open source and free if you run it yourself.
What's free: the software itself, with unlimited workflows. On your own machine it runs without spending a cent; to make it permanently reachable, you need a small rented server starting at about 5 euros a month.
The limits, and this is where it often gets dishonest: First, "free" here doesn't mean "effortless." You install, update, and repair it yourself; this is a path for techies. Second, the AI block in the workflow needs a language model, and using one via API keys costs extra, billed by consumption. So the agent is only completely free as long as you also run a local model, and that's another level of tech on top. Third, n8n's cloud has no permanent free plan, only a 14-day trial; after that, the Starter plan begins at 20 euros a month, billed annually (as of August 2026).
More fundamentally: an n8n workflow is an assembly line, not a worker with judgment. Which tasks that fits, and why my own agents still don't run on workflow platforms, is covered in Building AI agents without code: my way.
The three paths compared
| Claude Free | ChatGPT Free | n8n self-hosted | |
|---|---|---|---|
| Real agent with its own knowledge | yes, via Projects | no, new custom GPTs are business plans only | yes, as a workflow |
| Technical hurdle | none | none | high (server, updates, API keys) |
| Running costs | 0 | 0 | from about 5 euros/month for a server, plus model API |
| Biggest limit | tight message allowance | no permanent context | maintenance and assembly-line logic |
When paying pays off
My rule of thumb: free is for finding out, not for working. If your agent saves you even one hour a week during the trial phase, a 20-dollar subscription has paid for itself in the first month, measured against any realistic hourly rate. The detailed math on what AI employees cost per month and what they replace is in AI employee costs.
What you shouldn't pay for: expensive agent builders and courses before you even know which task your agent is supposed to take over. The task first, then the tool, and only then the money.
Where to start
Start today with Claude Free and a single real task: create a project, write instructions, upload three examples, test. Costs nothing but an hour, and afterward you'll know from your own experience whether the subscription is worth it for you. The step-by-step guide is in Build an AI agent with Claude.
