Terminal-based coding agents have become a genuine category rather than a novelty, and the interesting question about this one is not whether it works, because it does, but where it sits against alternatives that got there first. Let me deal with the strongest argument for it immediately, which is the price. There is a free tier that requires nothing but signing in with a personal Google account, and the daily allowance is large enough for real work rather than a token trial. For a student, a hobbyist, someone in a country where the paid alternatives are genuinely expensive, or anyone who wants to find out what agentic coding is like before spending money, this removes the barrier entirely.
That is not a small contribution and it deserves to be said first. Being open source is the second substantive point. The whole thing is readable. If you have ever wondered how these agents actually work, what the system prompt looks like, how the tool definitions are structured, how the loop decides when it is finished, you can simply go and look.
That is educationally valuable in a way that a closed product cannot be, and it also means you can fork it, add your own tools, or change behaviour you do not like. Anyone building their own agent should read this codebase. The very large context window is a real differentiator on a specific class of task. Asking questions that require seeing a whole repository at once, understanding how a change ripples through many files, or summarising an unfamiliar codebase, are things this handles better than tools working with smaller windows and cleverer retrieval.
If your work involves a lot of that, it is a genuine advantage. Extensibility through MCP servers is handled properly, and the project configuration file for persistent instructions works as you would expect. The tooling ecosystem around these agents has converged on sensible conventions and this participates in them rather than inventing its own. Now the honest comparison, because pretending otherwise would not help anyone.
On the specific task of editing existing code across several files, this is behind the leading alternatives and the difference is noticeable in ordinary use. It applies edits that do not quite land. It occasionally loses track of a change it made two steps ago. On a multi-step refactor it is more likely to need correcting, and correcting an agent is the expensive part of using one.
For single-file work and for greenfield code the gap narrows considerably, and it is the messy multi-file work on an existing codebase that most professional development actually consists of. The scope creep is the behaviour that irritates me most. Ask for a specific change and you will regularly get that change plus reformatting, plus a refactor of an adjacent function, plus improved error handling you did not request. Some of those changes are improvements in isolation.
None of them belong in the diff you asked for, and reviewing a change that does four things when you asked for one is slower than doing it yourself. Instructions in the configuration file help and do not eliminate it. The pace of development cuts both ways. Improvements land quickly, which is good, and it also means documentation, blog posts and community advice go stale fast, and behaviour you relied on can change between versions.
Anyone writing about this tool is writing about a moving target and readers should treat any specific claim, including mine, as time-limited. The data handling question is the one users should think about and mostly will not. The free tier is free because usage may be used to improve the product, and the terms differ between the free tier and paid API access. Everything is documented and none of it is hidden, and it is also the kind of thing people click past.
If you are pointing this at a client's proprietary code or anything under a confidentiality obligation, read the terms properly and consider whether the paid path with different handling is the appropriate one. Free has a reason. My three point five is for a capable open source agent, an unusually generous free tier that genuinely lowers the barrier to this whole category, real strength on large-context whole-repository work, and proper extensibility. Marked down for multi-file editing reliability that lags the leaders, for a persistent habit of doing more than asked, and for documentation that struggles to keep up with its own release pace.
Install it, it costs nothing. Judge for yourself whether it replaces what you use now.