Twenty Agent Platforms: Open-Source Is Not Free Software
Twenty agent platforms on one card: ten you pay for, ten open-source ones you do not. It is a genuinely useful map of a confusing category – and the most important line is the smallest one, printed along the bottom: open-source is not free software. You still pay for models.
That sentence is the difference between a cheap experiment and a surprise bill, so it is worth unpacking before any of the names.
An agent has four bills, not one
| Bill 1 · The platform → | Bill 2 · The model → | Bill 3 · Hosting and upkeep → | Bill 4 · Your attention |
|---|---|---|---|
| A subscription, or nothing This is the only one the chart shows |
Tokens, on every run Charged either way, paid or open-source |
A server, updates, breakages Yours to carry if you self-host |
Reviewing what it did Every autonomous run needs a human read, at least at first |
| Paid: monthly. Open: zero | Both: per use | Paid: included. Open: yours | Both: yours |
The fourth bill is the one nobody prices. An agent that acts on your behalf produces work you have to read before trusting – especially in its first weeks, when you are still learning where it goes wrong. That review time is real, and it is why one well-scoped agent beats five ambitious ones.
The twenty, and who each is really for
| Platform | What the card says it is | Column | Who it is really for |
|---|---|---|---|
| Manus | General agent that does tasks | Paid | Curious individuals and small teams |
| Lindy AI | Business agents, no code | Paid | Businesses without developers |
| Relevance AI | Agent teams for companies | Paid | Operations teams |
| Agentforce | Salesforce’s agent platform | Paid | Companies already on that platform |
| Copilot | Agents across Microsoft 365 | Paid | Anyone already paying for the suite |
| Glean | Agents over company knowledge | Paid | Companies with a lot of internal documents |
| Superagent | Autonomous security agents | Paid | Security teams |
| Gumloop | Drag-and-drop work agents | Paid | Non-developers automating office work |
| SmythOS | Visual agent builder | Paid | Builders who prefer a canvas to code |
| OpenHands | Open-source coding agent | Open | Developers |
| CrewAI | Crews of role-based agents | Open | Developers building multi-agent systems |
| AutoGen / AG2 | Multi-agent chat framework | Open | Developers and researchers |
| LangGraph | Graph-based agent flows | Open | Developers who need control over the flow |
| Dify | Self-host agent workflows | Open | Teams who want to host their own |
| Flowise | Build agents visually | Open | Technical users who like a canvas |
| Google ADK | Google’s agent dev kit | Open | Developers on that stack |
| Agents SDK | OpenAI’s agent library | Open | Developers on that stack |
| Mastra | TypeScript agent framework | Open | JavaScript developers |
| Agno | Agent framework | Open | Developers |
Read the right-hand column and a pattern appears. The paid side is mostly enterprise software – agents over company knowledge, agents for sales teams, agents for security operations. The open side is mostly developer libraries and frameworks: things you install, host and debug.
If you are one person making videos, neither column is aimed at you. That is not a gap in your setup; it is the same “Best column” problem we described in the wall is the signal, not the chart.
The post this came from
Watch the reel
Ten agent platforms you pay for, ten open-source ones you do not. Posted on the BANI Academy page.
What an agent is, in one paragraph
The word is used for everything at the moment, so it is worth being precise. An assistant answers when you ask. An automation runs a fixed sequence you defined. An agent is given a goal and decides the steps itself – it plans, acts, checks the result and keeps going until it believes the goal is met.
That difference is the whole reason agents are exciting and the whole reason they are risky. Deciding the steps is what lets an agent handle a messy job you could not have written out in advance. It is also what makes its mistakes harder to predict than an automation, which fails in the same place every time.
The practical consequence: give agents goals that are checkable. “Rename these files by date” has an obvious right answer you can see at a glance. “Improve my channel’s positioning” does not, and an agent will produce something confident and unverifiable. Checkable goals first, always – then widen scope once you trust the pattern.
What a creator should build instead
Almost everything a one-person channel wants from “agents” is covered by two much smaller things: one connector-style automation, and an assistant with your standing instructions.
| The job | What a creator actually needs | Why | What not to build |
|---|---|---|---|
| Move information between apps | A no-code automation | Minutes to set up, no maintenance | A framework that needs hosting |
| Draft something from a template | A saved instruction in your assistant | Free, instant, predictable | An agent that rewrites your voice |
| Do a multi-step chore on files | A desktop or browser assistant, watched | Reversible, on copies | An autonomous loop you do not read |
| Answer viewer questions | You, or a draft you approve | Trust is the product | An agent replying unsupervised |
| Anything touching money | Nothing automated at all | The only safe setting | Any of it |
The split between the connector and the worker is the one we drew in two of these are the automation: the flow moves the work, the assistant does the thinking, and neither needs a framework you host yourself.
When something does act on your accounts, the rules are the same as ever: narrowest access, read-only first, watch the first runs, and treat anything it reads as data rather than instructions – stop pasting. Autonomy multiplies mistakes as efficiently as it multiplies work, which is why the review column exists in the first place, as in same brain, three doors.
Why agent costs surprise people
A normal request costs one model call. An agent is a loop: it plans, acts, reads the result, re-plans, and repeats until it thinks it is done. A job you imagined as one call can easily be thirty, and a stuck agent can retry far more than that before anyone notices.
Three habits keep that in check. Set a limit on how many steps a run may take, if the tool allows it. Test on a small input first, and look at the run log rather than just the output. And check the bill after the first fortnight, not at the end of the month – the same discipline as the retry loop in you are paying for retries.
It is also worth deciding in advance which currency you would rather spend. A paid platform buys setup time; a self-hosted framework buys control. That trade is the subject of money or time, and agents are the clearest example of it in the whole tool landscape.
Frequently asked questions
Is open-source cheaper?
Sometimes. You save the platform fee and take on hosting, updates and debugging – plus the same model costs as everyone else.
Do I need to know how to code?
For most of the open-source column, yes. The paid column is built for people who do not, which is what the subscription buys.
Why do agents get expensive?
Because they loop. A single request is cheap; an agent that retries, re-reads and re-plans can make dozens of calls before it finishes.
What should I automate first?
The chore you have written down three weeks running – and let it produce a draft, not a published thing.
Price the run before you build the agent
Take the chore you were about to automate. Do it once by hand, note how many separate steps and how much text is involved, then ask what that would cost in model calls if it ran daily. If the answer is uncomfortable, the fix is a smaller scope rather than a different platform.
Open-source is not free software; it is free of a subscription. The models, the hosting and the reviewing are yours either way – so start with one narrow automation, keep it read-only, and let it earn the next one.
Learning a lot does not make anyone good. Doing a lot does.
If you want the structure that turns that into a habit, that is what our free challenge is built for: no barrier to entry, a genuine commitment to action, and a refundable commitment fee you get back when you finish. You are not paying for the knowledge. You are betting on yourself finishing, and we hold the stake.
Which agent platform would you try first? Tell us in the comments on the original reel.
For more on building channels for English-speaking audiences, that is what we work on at mmoyoutube.com.
