Published: 2026-09-15
Summary

Freebuff: an ad-funded coding agent with a daily allowance of GLM 5.3 Flash, Luna and DeepSeek

Chapters / key moments (click to jump — plays here on the page)

A real option for anyone priced out of frontier subscriptions. Freebuff is funded by ads shown in the IDE, not by a card. It is not a frontier setup. The point is that capable open-weights and small models are now good enough to build with, and this removes the payment barrier. He is not affiliated.

Source video

"I Found a Way To Use AI Agents Like Codex Completely For FREE" by Nick Saraev — Watch on YouTube →

What you get

Pricing model: a daily allowance of "free bucks" that refills at midnight Pacific. An hour of any model is a fixed price charged at session start. There are no weekly or monthly caps, only the daily allowance.
Models shown: GLM 5.3 Flash (about 20 hours/day), GPT-5.6 Luna (about 5 hours), DeepSeek V4.1 Flash (about 6 hours), and Muse Spark 1.2, among others.
App features: build and plan modes, several tabs running in parallel that can talk to each other, a diff/files/preview/terminal side panel, select-an-element-and-instruct, project settings including an agents.md, mirroring to the iOS app, and a "Merge PR" mission (finish the task, open a PR, drive checks to green, merge).
Ads: shown in the side panel while tokens stream. In his test they were not intrusive.

Gotchas & Caveats

The macOS download was an unverified binary (he says so on screen). Check the publisher and signature before running it. That is our standard advice for any agent binary. See security.
Heavy local CPU and fan use during his session. The reason wasn't explained.
Output quality matched the models: a working but plain website and a simple game. The free tier won't reach Fable 5.1 or Astra quality.
Pay attention to the "Merge PR" mission. An agent that can merge its own PRs needs branch protection set up before you enable it.
Ad-funded pricing may not last. Don't build a workflow that depends on it.

Key Takeaways

It's a way to learn agentic coding without a card, on models that score within reach of the frontier on coding benchmarks.
Use the plan mode and agents.md habits here as you would in a paid harness. They carry over.

Other low-cost routes: best free models for Hermes · cost optimisation.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Install and sign in
He downloaded the macOS app (an unverified-developer warning appeared), opened it and signed in with Google; no credit card is required. Projects open in a window with a chat, a diff view, a file list, a preview and a terminal.
Spend your daily 'free bucks'
A fresh daily allowance refills at midnight Pacific and can be spent on any model; the cost of an hour of any model is one price charged once when the session starts, with no weekly or monthly session caps. The site lists free time on models such as GPT 5.6 Luna (about 5 hours), DeepSeek V4.1 Flash (about 6 hours), Muse Spark 1.2 and GLM 5.3 Flash (up to about 20 hours).
Use the controls
Switch between build and plan mode, select an element on the preview to attach to your message, open a project settings panel and add your own AGENTS.md, run a 'mission' such as merge-PR (open a pull request, drive checks green and merge), and mirror the session to a phone through an iOS app.
Run two sessions at once
He opened a second project in another tab with a different model (GLM 5.3 Flash) at the same time; the two threads can talk to each other.

Gotchas

It is funded by ads shown beside your session, for example a pet-food advertisement; he found them unobtrusive and says he is not sponsored.
His computer got hot and loud during the first run, and the first result took over a minute.
Quality is that of smaller models; his generated site and stick-figure game were basic.
The economics are untested; the speaker says we'll see how the business model holds up.