Not Only Monthly Subscriptions for Development with AI Agents

How I got here after trying the available tools, and what I settled on — including the free bonuses

App Screenshot

I remember clearly the times when a chat with a neural network could not do even simple things in Android development. For example: lay out a screen so that if the items in a list are not empty, the list itself is displayed, and if they are empty, some placeholder text is placed in the centre of the screen with a button underneath. But technology decided to advance at an unheard-of pace, and the time when ignoring vibe coding and its tools had any rational justification is long gone. Today, vibe coding itself is associated for many with two flagship solutions that only the lazy have not heard of.

Anthropic and OpenAI offer their tools both for the terminal and for the GUI. VSCode extensions and a separate VSCode-based IDE, Cursor, seem to get a little less attention. At the same time, it seems to me that many people are put off by the subscription system, even at a price starting at around $20 for a basic subscription (there are exceptions and a price range, but monthly plans still seem to land around that amount). And how do you try a subscription system when you may be taking a risk, since your country of residence may ban the use of Claude Code and Codex, and few people want to take a risk and hand over money with the threat of an account ban hanging over them (and precedents already number in the thousands).

Despite the various optimisations and advantages that Claude Code and Codex undoubtedly have, their downsides must be mentioned too. First, there is the lock-in to their ecosystem. As a result, you cannot properly test models from other companies (there are ways, of course, but they look like a potential loophole that will be closed in the future). Second, on social media I have often seen that people feel pressure from the subscription period, adjust their lives to daily and weekly windows, and this creates additional anxiety and haste to load the agent while they still can.

Learn more about Claude Code

Learn more about Codex

What Can You Try Completely for Free, and How Far Does It Go?

Before buying a subscription, you always want to first understand what the proposed ecosystem and tool are like.

Cursor and Antigravity

You can try them for free before paying, but the limit is very small. I had the same task of fixing Preview in a KMP project when using Lottie. Both ADEs (Agentic Development Environments) and their models handled the task almost equally well and found a solution. In my opinion, Google did slightly better (it is almost the same as what Cursor produced on a Grok model, just a bit prettier). But the main thing is that the limits were enough for exactly one task. And it is not the first time I have had the feeling that even Google’s Flash models work somewhat slowly; I had the same feeling earlier when I used the chat-with-agent mode directly in Android Studio, although some benchmarks indicate that the models cope quickly with many tasks, which I cannot confirm from my subjective experience.

Learn more about Cursor

Learn more about Google Antigravity

Codex

From my observations, OpenAI offers the most generous limit without any subscription at all, albeit with models that are not the most powerful. And the most surprising thing is that after a month the limit resets, which is also a plus. I feel it will not always be like this, and it seems that sooner or later the company will close such generosity, but for now the opportunity exists and it is pleasant. And what is even more surprising: in my personal free version, there are only monthly limits. Weekly windows were not displayed at all, even in the settings, and most likely they are not even taken into account.

Learn more about Codex

OpenCode

OpenCode offers both CLI and GUI solutions. Moreover, personally I have not even tried CLI solutions in principle yet. I liked the format of a GUI agent with chats and an IDE open next to it so much that I got the feeling that when you are working on your own projects and need to experiment a lot, having an IDE open next to it is exactly what you need. And sometimes, say, for a minimal layout tweak, you do not even want to hand it to the agent. You want to visually move the View yourself and understand what is better. And OpenCode has the cheapest subscription — at the time of writing, it is $10, which is also a discount for those who want to try it first.

But what is far more interesting is that OpenCode often provides a list of free models and is not tied to a specific vendor. It is interesting to follow OpenCode and try something new.

Learn more about OpenCode

We Pay for Tokens, Not for a Month — and with Free Bonuses

OpenRouter

This is a model provider operating on a “Pay-as-you-go” basis. After registration, a key is generated that can be used in many places: from VSCode or Cursor extensions to ADEs or GUIs like OpenCode. And for me right now, my personal top is exactly the combination of OpenCode and OpenRouter. And most importantly, despite the fact that with an API it is logical that usage will be more expensive, in return you get flexibility: you do not have to use the most powerful and expensive models for every single task. A model should be chosen according to the complexity of the task. And thanks to the availability of choice — including models that are periodically given away for free — you can quite easily vibe code for less than $20 and experiment at the same time, since there is no limitation or lock-in to an ecosystem and no tension from windows and subscriptions, as with Claude Code or Codex.

And another important point to note: OpenRouter offers models from OpenAI and Anthropic.

But since OpenRouter is only a provider, I will briefly go over my experience of what to pair an OpenRouter key with.

Learn more about OpenRouter

Candidates to Pair with OpenRouter, Besides OpenCode

I will mention the GUIs I installed on my laptop out of curiosity, but which in my subjective opinion deserve attention or watching in the future.

ThinkRail

A GUI from JetBrains. It does not consume much RAM. In terms of design, in my opinion, it looks the nicest of all, but for now it puts me off because you cannot specify a file for context in the chat via @. Or maybe you can and they have a different invocation pattern and I have not figured it out, but for now that is how it is. So I will just keep an eye on updates, since it is an interesting candidate for a full switch in the future.

Learn more about ThinkRail

Zed

Lightweight. It lets you edit code and specialises more as an ADE than just a GUI for managing chats with agents. And when you first open a project, it immediately offers to pull in the necessary plugins for syntax highlighting.

Learn more about Zed

Takeaways

Vibe coding can be built without limiting yourself to a specific ecosystem, which gives good flexibility and room for experimentation, which as a result can also help you save money (no guarantee, it all depends on choosing the right models for the task). And equally important, it helps remove the inconvenience of trying to fit into daily or weekly windows; you feel clearly more free, and for me this is not a question of some discipline but a question of convenience.

In this note I have listed far from all the tools you can try, but it seemed that these are some of the most popular ones outside the ecosystems of the giants that have conquered the market.