Freebuff Tested for Three Days: What the Free Coding Agent Really Costs

A coding agent that offers DeepSeek V4.1 Flash for $0 sounds too good, and in a sense it is. Freebuff really is free to start, but the price gets paid in other currencies. A three-day trial of the desktop app showed a daily usage meter, ads and, for one model, a data-training label. Here's how each works, so you can decide in five minutes rather than three days.

What Freebuff is

Ads pay for Freebuff, a coding agent made by a small Y Combinator group. Its base is Codebuff, the team's open multi-agent framework, and the code sits in a public GitHub repo with an Apache-2.0 license. It comes as five products:

• Desktop: run several agents in parallel, each in its own workspace, on macOS, Windows or Linux
• CLI: a coding agent in the terminal
• Web: a prompt-to-deployed-app builder in the style of Lovable
• Cloud: an agent in a cloud IDE tied to your GitHub repo
• Chat: a research-style AI chat

The homepage makes its pitch bluntly, with a price chart that puts Freebuff at $0 a year beside Cursor, Codex and Claude Code. There's no API key and no credit card. The terminal version installs with one line:

npm install -g freebuff
cd your-project
freebuff

The three days here went into Freebuff Desktop, pointed at a Swift side project.

A different model menu

The model list was the first surprise. On the website six are named, among them GPT-5.6 Luna and Meta's Muse Spark 1.2, yet the in-app picker offered just four, grouped under a grey Limited access tag with a short note underneath.

The README has the reason. Users in supported regions see everything. People elsewhere, or connected through a VPN, are limited to a list of four: Solar Pro 4, MiMo 2.5, DeepSeek V4.1 Flash and GLM 5.3 Flash. The picker looked like this:

MODEL                 LABEL                          COST
GLM 5.3 Flash         Recommended · Deep reasoning    5 / hr
MiMo 2.5              Balanced                       10 / hr
Solar Pro 4           Limited-time trial             10 / hr
DeepSeek V4.1 Flash   Smart & Fast                   15 / hr
                      "May use data for AI training"

Pay attention to the last column, because the first catch lives there.

Catch one: the Freebucks meter

Every model has an hourly price in Freebucks, and a single daily allowance covers all of them. The homepage promises 100 Freebucks a day with full access, and on limited access the picker showed 25. Now do the math for the model that attracted attention in the first place:

25 Freebucks ÷ 15 per hour  =  ~1 hour 40 minutes of DeepSeek V4.1 Flash
25 Freebucks ÷ 5 per hour   =  ~5 hours of GLM 5.3 Flash

DeepSeek V4.1 Flash is the most expensive entry on the menu, and a whole day's allowance of it is shorter than a film.

The limit shows up in practice. At 4:02 PM, the prompt "run app and test it" produced no code. It produced a notice saying Freebucks would return in 7 hours 57 minutes. A /preview attempt four minutes later hit the same wall. That message also points to a way out: more Freebucks per day starting at $8 a month, and the DeepSeek row has its own "7x usage for $5" upsell. None of that is shady, since somebody has to pay for GPU time. It does mean you should choose models as though you're spending money, because you are.

A better plan is to begin on GLM 5.3 Flash. It's preselected, tagged for deep reasoning and a third of DeepSeek's price. Keep DeepSeek V4.1 Flash for short stretches when speed is the point. One more detail from the README: models may be served from a quantized (Q8_0) build, so if one behaves slightly differently from the provider's own API, that could explain it.

Catch two: ads

Asked how it can be free, the FAQ says just: text ads. That's an acceptable deal for a no-cost tool, and ads are better than an unexpected invoice. Read the part people skip, though. According to the README, Freebuff may analyze your prompts and messages to personalize those ads. It also says separately uploaded files and connected repositories aren't passed to advertising providers. So your code doesn't go to advertisers, but what you type to the agent can influence what you see. If your prompts describe a client, an unreleased product or anything under NDA, bear that in mind.

Catch three: the training label

Go back to the DeepSeek V4.1 Flash row. Below "Smart & Fast" a second line reads May use data for AI training. That label has real meaning. The README says that when a model carries the label, Freebuff or the model provider may retain what you submit to train, evaluate and improve AI models, and restricted partners may evaluate code used with those models.

Think of an iOS codebase for an investment app. That's precisely what you shouldn't paste into a model with a training label. A workable rule:

• Hobby projects, open-source contributions, practice repos: use whichever model you like
• Employer code, customer code, anything containing secrets: avoid training-labeled models, and check your employer's policy before using a free tier at all

Three days of actual use

With the catches covered, here's what justified calling it good. Desktop doesn't feel like a free toy. It's a real agent workspace, with chat on the left and tabs on the right for Queue, Changes, Files, Preview and Terminal. A Build/Plan switch beside the prompt box lets the agent think before it touches files.

A loose, natural-sounding request for the Swift project, along the lines of asking it to analyse the tool in depth and improve its drawing accuracy, got more than advice back. A new test file, Tests/FaceRenderProbe.swift, appeared with 111 added lines, and the count in the Changes tab reached 213. That behavior belongs to an agent operating inside a repo, and a chatbot pasting snippets wouldn't do it.

A few other details stood out:

• Launch preview: the agent can open an HTML page or start the project's dev server and show it inside the app
• Connect a provider: enter your own API keys, pay that provider directly, and carry on in the same conversation
• Existing agents: according to the README, Desktop can also drive a Claude Code or Codex that's installed locally and signed in to your account
• Young software: version 0.0.114 updated itself to 0.0.115 in the middle of a session, so rough edges are to be expected at this stage

Who it suits

It fits well if you're a student or junior developer who can't justify a $20 to $200 monthly subscription, if you build side projects or open-source work, if you want to try open models like GLM, MiMo and DeepSeek in a proper agent harness without API keys, or if you're in a full-access region where 100 Freebucks a day goes much further.

Think twice if you work on proprietary or regulated code, if a limited-access country is your base and you rely on long unbroken agent sessions daily, or if you can't stand ads in your tools.

So was it free?

The answer is yes. Nothing was paid for three days with a working coding agent that handled a real repo. But the homepage skips three costs: a quota that drains quickest on the very model that got you interested, ads that learn from your prompts, and a training label that determines which code you can safely use.