How to run AI coding agents like Codex for free by watching ads instead of paying

Every AI coding agent worth using costs money. Claude Code runs about $20 a month minimum. Cursor charges monthly. Codex needs a ChatGPT Plus subscription or an API key with a credit card attached. Even the "cheap" options add up fast once you're running agents daily.

So here's a strange but very real workaround: a platform that gives you access to AI coding agents completely free, no credit card, no subscription, no API key — because it's funded by ads instead of your wallet.

It's called Freebuff, and the pricing math it puts on its own homepage is almost aggressive. Where Codex costs $240 a year, ChatGPT Plus costs $240, Cursor costs $720, and Claude Code costs $1,200, Freebuff lists itself at $0 — forever. That's not a typo. That's the pitch.

What is Freebuff and how does it actually work?

Freebuff is a desktop and iOS platform that gives people free access to AI coding agents by showing ads instead of charging a subscription or requiring a credit card. Instead of paying per token or per month like you would with most agentic coding tools, you sign in, pick a model, and watch ads appear alongside your coding session while the agent works.

The primary way to access it is at freebuff.com, and the project is open source on GitHub under CodebuffAI/freebuff. That matters, because Freebuff isn't some random side project — it's a terminal AI coding agent built by James Grugett, co-founder of Codebuff (a Y Combinator company) and Manifold Markets, that launched in March 2026. The team behind it already had experience building Codebuff, a subscription-based coding agent, before flipping the model entirely to ad-supported and free.

The credits system is straightforward. The platform runs on a daily allowance system: free credits refresh at midnight Pacific, and you spend them on whichever model you want, with pricier models burning through the allowance faster than cheaper ones. So a lightweight model might last you all day, while a more expensive one chews through your allowance in twenty minutes.

Why can AI agents suddenly be free?

This isn't charity. It's economics catching up with reality. Inference costs dropped roughly 10x in the past year, and models like DeepSeek V4, Kimi K2, and MiniMax M3 brought coding-grade performance to under $1 per million tokens — a price point that makes a flat $20/month subscription look like pure markup rather than a fair service fee.

Meanwhile Claude Code, Cursor, and Codex held their subscription pricing flat even as the underlying cost of running these models collapsed. That gap between what inference actually costs and what subscriptions charge is exactly the gap Freebuff is stepping into — building an ad-supported CLI where developer-relevant ads perform well enough to cover infrastructure costs, then giving the actual coding experience away.

It's worth noting this isn't the first attempt at this model. A previous ad-supported coding tool called AMP Code lasted a few months before that approach wasn't sustainable, but Freebuff has been running for several months now and seems to be holding up better. Whether it lasts long-term is genuinely an open question — but right now, it works.

Which models does Freebuff actually give you?

This is where it gets interesting, because the model lineup isn't static. The platform offers hourly access to multiple models, including GLM 5.3 Flash, DeepSeek V4.1 Flash, and a GPT-branded model referred to as GPT 5.6 Luna, each with different daily hour allotments.

Some models are essentially unmetered. GLM 5.3 Flash, DeepSeek V4.1 Flash, MiMo 2.5, and Solar Pro 4 are unmetered at full access and cost no session at all — meaning you can hammer away at those without watching your daily allowance drain. Others draw from your shared pool.

None of these are frontier-tier reasoning models on the level of GPT-5.4 or Claude Opus. The models on offer aren't frontier-tier, but they're capable enough for practical coding tasks like scaffolding a site, writing simple scripts, or building small games. That's an important expectation to set before you dive in — this is a tool for prototyping, learning, and small projects, not for shipping mission-critical production infrastructure.

How do you actually get started with Freebuff?

Getting set up takes less time than reading this section. Head to freebuff.com and follow this flow:

1. Sign in with Google — there's no email/password form, no credit card field, and no plan selection screen. You just authenticate and get dropped back into your terminal or browser session.
2. Install the CLI if you're using the terminal version. The Freebuff GitHub repo has installation instructions, and it's also published on npm — you can start using Freebuff in your project by running npm i freebuff.
3. Pick a model. Start with one of the unmetered options (GLM 5.3 Flash or DeepSeek V4.1 Flash) if you want to conserve your daily allowance for a heavier model later.
4. Describe what you want built. The workflow mirrors any other agentic coding tool — type a prompt, watch the agent reason and write files, preview the result, and iterate. Simple requests, like a one-page personal website or a basic browser game, generated working output within a couple of minutes, complete with visual preview and the ability to select page elements to target follow-up instructions.
5. Watch the ads, keep coding. The ad load during typical use is moderate, appearing as small inline units rather than forced video interruptions.

If you want the full-stack, browser-based version instead of the terminal CLI, there's also Freebuff Web, which lets you build, preview, and deploy full-stack apps for free with no subscription and no API keys required.

How does this compare to Codex, Claude Code, and Cursor?

If you already have a paid subscription to OpenAI Codex CLI, don't cancel it just yet. Freebuff is genuinely useful, but it has real limits. Freebuff won't replace Claude Code or Cursor for complex backend work — the models are capable for front-end tasks and straightforward coding, but they hit limits with more demanding projects.

Where it shines is as a learning tool and a budget safety net. It's a great fit for beginners who want to understand how AI coding agents work without spending money — you get a feel for the CLI workflow, learn how to prompt effectively, and see what these tools can do. It's also solid for developers on a budget who need help with small tasks — if you're building a personal project or need a quick prototype, Freebuff handles that fine.

If you're currently paying for OpenAI's tool, it's worth comparing directly. The official Codex documentation on developers.openai.com walks through installing the CLI locally and choosing models like GPT-5-Codex, and the Codex GitHub repo has the one-line install script if you want to run both tools side by side and see where each one earns its keep. Run your next few small projects through Freebuff first, and reach for a paid agent only when you hit a wall it can't climb.

What should you actually use Freebuff for?

Think of it less as a replacement for your main coding agent and more as a free lane for the work that doesn't need your best model. A few practical use cases:

• Prototyping ideas before committing budget. Test whether an app concept even works before you spend a cent on API credits.
• Learning agentic coding workflows. If you've never used a CLI-based AI agent, this is a zero-risk way to build the muscle memory.
• Small personal projects and scripts. Landing pages, browser games, utility scripts — tasks like scaffolding a site, writing simple scripts, or building small games are exactly what it's built for.
• Backup capacity when you've burned through paid credits. If you've hit your monthly cap on a paid tool, Freebuff can pick up the slack for lighter tasks.

Is watching ads for AI worth it?

Here's the honest trade-off: your attention and a bit of data usage in exchange for compute that would otherwise cost real money. Freebuff uses prompts, messages, agent traces, code, files, and repository data to provide the service, and may analyze prompts and messages to personalize ads. If that data arrangement doesn't sit right with you for sensitive or proprietary codebases, keep those projects on a paid, private tool and save Freebuff for the throwaway stuff.

But for the person who's been putting off learning AI-assisted coding because of subscription fatigue, this removes the last excuse. No credit card, no trial period that quietly converts to a charge, no API key setup. Just sign in and start building.

The inference cost curve isn't going back up. More tools like this are coming, whether the current subscription-based giants like it or not. The only question left is whether you start experimenting now, while it's still a novelty, or wait until everyone else already has a head start.