FreeBuff vs Clixad: two ad-funded coding agents, and where the ad sits

Two products that share a funding word

An earlier post here argued that "ad-supported" is doing too much work as a label, because three different arrangements are sold under it: in-session ads, a vendor-funded daily grant, and a rewarded offerwall. That post also concluded that the in-session model had been tried by the best-resourced company in the category and abandoned inside six months, when Sourcegraph dropped ads from Amp in March 2026.

That conclusion needs a correction, and this post is partly it. FreeBuff launched on 12 February 2026 with in-session text ads and is still running them in August. One vendor leaving a model does not close it, and a second implementation that has lasted six months is evidence the first read was too quick. So this is the concrete version of that earlier argument: two live products, both ad-funded, compared on what they actually grant.

What FreeBuff grants

FreeBuff is the free, ad-supported build of Codebuff, published under Apache-2.0. Everything in this section is from FreeBuff's own launch post and README, linked in Sources; where two review sites disagreed with each other about the session limit, I have used the primary source and ignored both.

• No account, no card. The launch post's words are "No account. No credit card. No config file." You install the CLI and start.
• Limited mode: "Deepseek v4 Flash only, 6 one-hour sessions per day." This is what users outside the supported regions get, and what VPN users get.
• Full mode: a wider model picker, described at launch as available in 25+ countries. The exact roster has moved between the launch post and the README, so check the current one rather than trusting a number in any article, including this one.
• The ads: "small, text-only ads from sponsors" between agent turns — not inside generated code, and not as popups.

That is a genuinely low-friction product. There is no signup, no wallet, no earning step and no moment where the tool asks you to go and do something else first. It is worth saying that clearly before the part where I disagree with it.

What Clixad grants

Clixad meters usage in credits against real token consumption. The rate is fixed in code: 100,000 credits is one US dollar of advertiser payout. A new account starts with 25,000 credits, and a free account can earn at most 500,000 credits in a day.

There is also one sponsored line above the input box while a model call runs, marked Ad ahead of the text. It is a trickle rather than a way to top up — a fraction of what completing an offer pays, and often nothing — and CLIXAD_SPONSOR=0 turns it off. Where the ad sits covers what it is chosen by and what it costs you, since that is the part worth arguing about.

Credits convert into work directly, which is the one advantage of metering something real. The table below prices the same twelve-call agent task used across this blog — one instruction that the agent turns into twelve model calls, resending the whole conversation each round — on every model in the catalogue, and then divides two Clixad grants by it.

The zeroes are the useful part. A median completed CPX offer does not buy one agent task on most of the catalogue, and the daily cap does not buy a single task on Claude Fable 5.1. Those rows stay in the table for the same reason the awkward numbers stay everywhere else on this site: a comparison that drops its own bad rows is not a comparison.

The other half of the honesty is the earning side. Credits arrive on completion, and the most common outcome of starting a survey is a screenout: a few questions in, it decides you are not the respondent it is paying for, and it ends with nothing credited. That is how offerwalls work industry-wide and I cannot engineer it away, because it happens on the wall's side before any postback reaches us. From our own ledger: completed offers on our CPX wall have credited from 9,000 to 114,000 credits on completion, with a median of 51,000: 26 completions recorded in our ledger between 27 July 2026 and 12 September 2026, as of 14 September 2026. The ledger does not record how long an offer took, so nothing here says what a minute pays.

The units do not convert

Here is where a comparison article normally produces a winner, and where this one cannot.

FreeBuff's free allowance is measured in time: six one-hour sessions a day in limited mode. Clixad's is measured in tokens, priced into credits. There is no exchange rate between them, and inventing one would be the exact error the tier-limits post exists to name.

What can be said without cheating: if your constraint is wall-clock time at the keyboard, a session limit is the thing that will bite you. If your constraint is how much the model has to read and write, a token meter is. Long sessions with little inference favour the meter. Short bursts of heavy inference favour the clock.

Where the ad sits, and what it has to see

This is the difference that does not come out in the wash, and it follows from geometry rather than from anybody's good intentions. It is no longer a difference about whether there is an ad on your screen. Both products put one there. It is about what is allowed to select it.

An ad rendered inside a coding session is selected by a process that is already holding your session. The vendor then has to decide how much of it informs the selection. FreeBuff answers that question in its launch post, and the sentence is worth quoting exactly rather than characterising:

"Prompts and messages may be analyzed to personalize ads, using Freebuff systems and service providers acting on our behalf."

The same page draws a boundary in the other direction — "Separate uploads and connected repositories are not provided to advertising providers" — and notes that models or features labelled "May use data for AI training" may keep submissions to develop and improve AI models. So by FreeBuff's own description: prompts and messages are in scope for ad personalisation, repositories and uploads are not handed to ad providers. That is a more careful line than most products in this space draw, and I would rather quote it than paraphrase it into something worse.

An ad that pays for a completed action in a separate tab has a different shape available to it. The advertiser is buying a finished survey or signup and pays a fixed amount for it; the payout arrives at our server as a signed postback naming a transaction and an amount. There is no step in that path where a prompt would help, and no step where one is present. That is not a promise about our character — a policy is a promise and can be revised — it is what the money path is made of.

Clixad's own in-session line is the case where that reasoning has to be applied to us rather than to somebody else. It is an advertisement on your screen while you work, so it belongs in this section and not in a footnote. What it is selected by is one fixed category, set on our server, which says the reader is using a developer tool in a terminal. The request carries a publisher id, that category, an opaque token standing in for your account, and a one-off value. There is no field in it a prompt could travel in. What it does cost you is a connection: it is fetched by your machine, so the ad network sees your IP address. CLIXAD_SPONSOR=0 stops the request being made at all, which is a switch FreeBuff's arrangement has no equivalent of, because its ads are the revenue rather than a top-up on it.

So the comparison narrows to something sharper than it started as. Two products, two ads on screen, and the question is whether the thing you are writing is an input to choosing them. FreeBuff's launch post says prompts and messages may be. Ours cannot be, because the selection never sees them — and the difference between "may be" and "cannot be" is the whole of what I would ask anyone to weigh here.

The trade is real in both directions. Clixad's arrangement asks you to leave the terminal, do something unrelated, and accept that it often pays nothing. FreeBuff's asks for none of that and instead lets the funding touch what you type. Which of those is worse depends entirely on what you are working on, and I do not think there is a general answer.

What each one is better at

Pick FreeBuff if you want to be coding in a minute with no account, no wallet and no interruption; if you would rather be shown a text ad than go and complete a survey; if a session-based clock fits your day better than a token meter; or if you want an Apache-2.0 codebase you can read.

Pick Clixad if you want frontier models on the same funding as the cheap ones, with no locked tier — a stronger model simply costs more credits per turn; if you would rather no part of your prompt be involved in selecting an advertisement; or if you want an allowance denominated in the thing you actually consume, so an unused hour costs nothing.

Pick neither if you already have an API key you are happy paying for. Both of these exist for people who do not, and a key you already hold beats both arrangements on predictability. That is not a difficult admission: the point of writing the numbers down is that they sometimes point away from you.

These two are also not the only options with inference included. The shortlist post puts both alongside Copilot Free, Cursor's Hobby plan, Amp Free and the Gemini CLI free tier, and applies the same test to all six: what does the grant actually finish?

FAQ