Freebuff: The Free Coding Agent Powered by Ads
In our piece comparing open-source coding agents, we talked about OpenCode and Pi Agent, among others—two free tools that require no login and never touch your credit card. The question left hanging was about frontier models, the ones that are truly expensive to serve. Can you offer a free coding agent without strict limits using the same models that power products costing hundreds of dollars a year? Freebuff tries to answer yes, with a bet that until recently would have seemed bizarre even in the developer tools world, famously allergic to any form of advertising: paying the bills with text ads slipped in between agent turns.
A Different Model
Freebuff is the product of CodebuffAI, a team that went through Y Combinator (batch F24) and is led by James Grugett, former founder of Manifold Markets. The project is structured across five products: Desktop, CLI, Web, Cloud, and Chat—all built on top of Codebuff, the open-source multi-agent framework (Apache 2.0 license) from which Freebuff inherits its architecture. The two taglines accompanying the launch, "we make coding 100% free" and "it's free because it should be", are not just marketing: in the first few weeks following its August 2026 debut, the product surpassed 230,000 developers, as Carbon Ads itself recounts in a profile dedicated to its most unusual publisher.
The economic mechanism is as simple as it is unusual for the sector. Between agent turns—never inside the generated code, never as invasive popups—text ads appear in the style of Carbon Ads, the ad network founded in 2010 specializing in technical audiences. The advertising is contextual rather than behavioral, following Carbon's historical model: if you are writing a backend, you are more likely to see an ad for a vector database than for a marketing automation tool, without this implying the creation of an individual developer profile. The underlying reasoning, as conveyed by CodebuffAI's materials, is that the cost of language models capable of writing code has fallen much faster than the price of the tools packaging them. Competitor subscriptions confirm this: OpenCode, which offers free models, starts at $120 a year for its paid tier, while Cursor, Claude Code, and Devin all sit around $240 a year for base plans, with premium tiers climbing well beyond. Freebuff costs zero, and for those who really do not want to see ads, there remains a paid escape hatch—Codebuff Pro—with plans at $100, $200, or $500 per month aimed more at teams and heavy usage than the casual developer.
Inside Freebuff
The architecture does not entrust everything to a single monolithic model; instead, it coordinates specialized agents: a File Picker Agent that explores the codebase, a Planner Agent that decides what to modify and in what order, an Editor Agent that writes the changes, and a Reviewer Agent that validates them. It is the same scheme as Codebuff, which in its internal benchmarks claims to beat Claude Code 61% to 53% on over 175 real coding tasks distributed across open-source repositories. These are biased numbers, to be clear, produced by the team selling the product, but they nevertheless give a measure of how central the multi-agent approach is to the offering.
The model catalog, at the time of writing, includes GLM 5.3 Flash for deep reasoning, DeepSeek V4 Flash 07/31 for fast writing and tool usage, GPT-5.6 Luna and MiMo 2.5 with image support, Solar Pro 4 featuring a 524,000-token context window (available only during a limited trial period), and Muse Spark 1.2, Meta's agentic model with a one-million-token context window, shared across the entire user base and therefore subject to queues during high-traffic periods. Access follows two parallel systems: the historical session-based system—six one-hour blocks per day extendable via referrals and bounties—and the newer Freebucks system, where every hour earns a small budget consumed in proportion to the cost of the chosen model. GLM 5.3 Flash, DeepSeek V4 Flash, MiMo 2.5, and Solar Pro 4 remain unmetered for those with full access, guaranteed in the United States, Canada, the United Kingdom, the European Union, and selected other countries; elsewhere, or behind a VPN, access is reduced to a subset of models.
In daily CLI usage, installable via npm install -g freebuff, the logic closely mirrors OpenCode and Pi Agent: you run the command in the project folder and describe what you need. One operational quirk is worth noting because it is not obvious: the timed session starts when it is created, not upon sending the first prompt, so opening it and leaving it idle still consumes minutes. Upon expiration, input locks, but the agent continues working in the background to complete the ongoing task, and the /history command allows retrieving the previously built context in a subsequent session so preparation work is not lost. Everything stays in local files, so nothing is truly lost, and the tool does not touch remote repositories unless explicitly asked. On the privacy front, CodebuffAI states that data is used to train models only when a feature explicitly indicates it, that prompts and messages may be analyzed to personalize ads, that separately uploaded files and connected repositories are not passed to advertisers, and that only a restricted number of partners may evaluate Cloud repositories or code associated with models labeled as usable for AI training.
GLM 5.3 Flash, the Ox Alpha Case
The model we chose to put to the test is GLM 5.3 Flash, and its recent story is worth telling because it says something about how open-weight models establish themselves today. On August 20, 2026, a model with no name and no declared company appeared on OpenRouter and OpenCode with a one-million-token context window. Within days, developers dubbed it Ox Alpha and put it through real tasks, from pull request audits to multi-agent workflows. In three days, it processed over 11 trillion tokens on OpenRouter alone—the most massive launch ever tracked by the platform—ending DeepSeek's 56-day reign at the top of the OpenCode leaderboards. On August 26, Zhipu AI (operating internationally under the Z.ai brand) confirmed that Ox Alpha was actually GLM 5.3 Flash, simultaneously releasing the weights on Hugging Face under the MIT license. Patrick Collison, co-founder of Stripe (which had agreed to acquire OpenRouter just the day before), commented on the story, describing it as "very impressive."
Technically, it is a Mixture-of-Experts model with 320 billion total parameters, of which only 18 billion are active per inference. It is the first natively multimodal model in the GLM-5 family, capable of accepting text, images, video, and files. The hybrid architecture, combining linear and sparse attention, reduces compute for attention by roughly three times and the KV cache by over four times compared to its elder sibling GLM-5.3, allowing it to sustain a one-million-token context without blowing up serving costs. Training took place on 30 trillion multimodal tokens. On the benchmark front, Artificial Analysis assigns GLM 5.3 Flash an Intelligence Index score of 57 versus 60 for the larger GLM-5.3, but at a cost per task of roughly $0.09 versus $0.68—roughly one-seventh. On Terminal-Bench 2.1, the most relevant benchmark for a coding agent, the score is 84.3, and on DeepSWE v1.1 it reaches 63.4. Zhipu also stated that Ox Alpha's entire stealth traffic week—over 62 trillion tokens in total—was served on a cluster of roughly 100,000 Chinese-produced AI chips with no Nvidia hardware involved, a statement that sent Zhipu shares in Hong Kong soaring over 12% and which, if confirmed at production scale, shrinks part of the infrastructure advantage that US export controls seek to protect.
AlboUp, Field Testing
To understand what all this means in practice, we built a real application with Freebuff and GLM 5.3 Flash: AlboUp, a web app to read the Official Notice Board (Albo Pretorio) of the Municipality of Verbania more easily than through the institutional portal. It is the third chapter in a personal suite of "*Up" apps dedicated to my local area, inspired by previous projects BenzUp and JobUp, written in pure HTML, CSS, and JavaScript with zero dependencies.
The app shows an immediate list of notices with type, summary subject, number, date, attachments, and a flag for new documents, featuring a detail view linking back to the official page. There are three sections—In Publication, Archive, and Marriages—with full-archive search, quick filters by notice type, progressive loading in 50-item pages, light and dark themes, local caching, and direct links to every single notice. It is also an installable PWA with manifest, icon, and animated splash screen.
The interesting part for anyone working with these tools lies in the technical challenges encountered along the way, as they truly reveal the model's level. The Municipality's portal, built on Liferay with the jCityGov portlet, exposes no APIs or CORS headers, so the solution identified by the agent went through the public proxy r.jina.ai, with a fallback and backup Markdown parser in case the service fails to respond—verified endpoint by endpoint without requiring sessions or cookies, covered by tests on real data. Another non-trivial detail involves digitally signed .p7m attachments, handled alongside regular PDFs, and the decision to use the notice publication period as the reference date instead of dates appearing within the body text, which are often misleading. Deployment is static, designed for Netlify, GitHub Pages, or Cloudflare Pages, with a test suite separating an offline parser.test.js from an e2e.test.js that checks online behavior where the most concrete bottleneck remains r.jina.ai's rate limit, mitigated by local caching.
In terms of timeline, the first substantially working draft arrived in one hour and twenty minutes, followed by about an hour spread over multiple sessions to refine details according to personal preferences. The strongest impression, beyond pure speed, was that of a model that systematically explores alternatives before choosing and then executes without excessive hesitation—a behavior resembling less a reactive code generator and more a colleague who first understands the problem and then solves it, with the economy of motion of an indie level designer who knows every corner of their map before even drawing it.
Can Ads Support Coding?
Freebuff's core bet is that developer attention is valuable enough to the right advertisers to fund a free product over the long haul. It is a thesis built on two concrete pillars: increasingly capable and cheap-to-serve open-weight models, as demonstrated by the story of GLM 5.3 Flash, and a niche advertising network like Carbon Ads that has worked with technical audiences for years without resorting to cookies or behavioral profiling. In the third quarter of 2026, the three most popular command-line agents are all free—a signal that the market is moving in this direction regardless of any single product. The most evident risk is dependency on the ad market, which can vary significantly by region, giving rise to the two access modes—full and limited—that characterize Freebuff today.
Viewed from another angle, Freebuff proposes a third path between the two existing options: not self-hosting, which requires technical skills and hardware beyond everyone's reach, and not paid API access, which still requires a dedicated budget. Advertising in this model demands neither infrastructure nor direct expenditure from the end user, theoretically democratizing access to frontier-model-assisted development tools. For developers, this means a genuinely free option in exchange for an explicit compromise on advertising and data usage for ad personalization—a compromise each person can weigh differently depending on how much value they place on their attention while writing code. For industry observers, it is a case study in how advertising can fund tools aimed at a technical audience historically most intolerant of any form of advertising.
Conclusions
Our experience with AlboUp shows that the experiment, at least in its initial phase, works: a free frontier model used to build a real application in a couple of total hours. GLM 5.3 Flash proves to be a capable model at a near-negligible serving cost, and its combination with Freebuff's infrastructure makes all this accessible without a credit card. However, more open questions remain than certainties. How sustainable over time is a model that ties free product quality to the advertising investment cycle in the tech sector, which is historically volatile? What happens when frontier model inference costs stop dropping at the current pace, or when competition among niche ad networks grows tighter? And above all: will the limited access mode reserved for entire geographic regions remain a temporary compromise, or will it become a structural fault line between those who have full access to assisted development tools and those who do not? These are not rhetorical questions, and answers will come only by watching what happens to Freebuff—and those who attempt to imitate it—over the coming quarters.