Freebuff: Why I pay $0 for coding agents on my private projects
This is a shameless plug — and it's still true.
I don't write my own code anymore. I've argued that at length elsewhere — for my private projects I let agents work and I stay the director. What I left open there: with what.
The honest answer is unglamorous. For my side projects I pay $0. Not "cheap", not "on the free tier" — nothing at all. The tool is called Freebuff (https://freebuff.com), and it is exactly what it sounds like: a coding agent for the terminal, funded by text ads.
What Freebuff is
Short and without the marketing speak: a coding agent that comes in several flavors. The one that matters most to me is the CLI — install via npm, change into a project directory, start it, describe what you want. No API key, no credit card - just an account. Under the hood runs Codebuff's multi-agent architecture: specialists for file finding, planning, editing and review instead of a single model.
There is also a desktop app (macOS, Windows, Linux), plus Web, Cloud (agents on any GitHub repo) and Chat — all funded by text ads. In the EU, Freebuff runs in "full mode" with the full model catalog (DeepSeek, GLM, MiMo and friends); elsewhere there is a "limited mode" with fewer sessions per day.
Why it is enough for my private projects
I almost never use parallel agents. I want to stay in the loop — to see what happens, to be able to intervene, to be the one doing the thinking. A swarm of agents taking the work off my hands while I look at another tab is not my model.
Yes, Freebuff is sometimes slower than the paid providers. I genuinely don't care: while the agent is working, I do something else. The bottleneck of my work was never the speed of the tool — it was my own thinking.
And I don't need SOTA flagship models either. Quite the opposite — they often want to take too much thinking away from you. I don't want the model making the architecture decision for me; I want to make it myself, with the agent as the executing hand and sparring partner. For my private stuff — [Kodama, eac-d, the little Go+SQLite site behind fratte.dev, a handful of throwaway scripts] — that has been more than enough so far.
If I had unlimited tokens, I would honestly use only GrokBot — it already worked successfully for one project. I don't, though. So the math is simple: roughly $120 to $2,400 a year for a hobby versus $0. For a tool that gets the job done day to day.
A note on GrokBot, because "several agents" is easily misunderstood: my experiment wasn't about many agents hammering the same code at once. What was interesting was the synchronization — the agents paced themselves like a small team. The architect reviewed a PR while the E2E bots were already testing and reporting their findings back to the builder, and so on. Not parallel typing, then, but a well-timed flow with clear roles and hand-offs. That is exactly the part that convinced me — and exactly the part I still want to direct myself.
Where the limits are
I won't sugarcoat this — Freebuff has clear trade-offs, and they belong on the table:
- Ads. The price is that ads show up in the terminal. If you can't stand that, pay elsewhere.
- Data. For some models it is explicitly stated that inputs may be used to improve the models. For code I own, or code I don't want to share, that's a deliberate no.
- Not the spearhead. The proprietary top models from the big vendors stay behind their paid plans. For most work that's irrelevant — for the last few percent I sometimes still reach for something stronger.
- Dependency. Freebuff is ad-funded, and therefore a business model that can change. I don't plan my existence around it.
Conclusion
For professional projects with budget and accountability, I keep using the best tool I can get. For my private projects: the most expensive agent is the one I don't need. Freebuff gives me the agentic way of working I've gotten used to — for $0, and so far without painfully missing the difference from a paid agent.
Little bling-bling, lots of use. That fits.