Codex · Claude Code · Cursor

The AI coding agent handbook, without the marketing

Most pages about AI coding agents are written from press releases. The ones that matter are written after the tool has failed on you at 11pm. That is what we collect here: the numbers, the failure modes, and the sequence that gets you working again.

Start here

Coming next

• queued Claude Code pricing: the real cost of a month
• queued Claude Code API key, without leaking it
• queued Codex not working — the full failure index
• queued Claude Code vs Cursor, by workflow instead of score

What an AI coding agent actually does

An AI coding agent is not autocomplete with better manners. It reads your repository, runs commands, reads the output, and decides what to do next — often for several minutes without asking you anything. That difference explains almost every complaint you will read about them: they cost more than a chat subscription, they fail in ways a chat window never does, and their errors are environment problems (paths, permissions, proxies) rather than model problems.

So the useful questions about an AI coding agent are practical ones. Which plan includes it, and what happens when the usage window closes? Which layer of the install broke when it will not start? Which tasks should you not give it at all? This site answers those with pages you can act on, not with rankings or reviews.

Why we write about failures instead of features

Feature lists are copied from press releases within a day, and they are the same everywhere. Failure knowledge is the opposite: it is expensive to produce, it ages slowly, and it is exactly what a search engine cannot answer because no single page owns it. When someone searches for an AI coding agent problem, they do not need another overview — they need the one page that tells them which of three causes they are hitting.

Who this is for

Developers who have already picked an AI coding agent and now need the unglamorous parts to work: the plan that does not run out at 3pm, the install that survives a corporate laptop, the error that comes back after an update. If you are still choosing between agents, the honest starting point is cost and failure modes, not a feature table — those are the two things that decide whether you keep using one after the first week.

Every page here is written to be finished with. Nothing is behind a newsletter, and there is no “part 2” that only exists to capture an email. If a page cannot answer your problem, that is a gap worth reporting, and it moves to the front of the queue.

Sources we link to

We prefer primary sources over summaries. The two we use most: Claude Code documentation for anything Anthropic ships, and the openai/codex issue tracker for what actually breaks in the field.

How we decide what to write next

Two inputs decide the queue. The first is search data: we check a keyword's difficulty against the actual top ten and only write a page when a small site can plausibly take a slot. The second is failure frequency: when the same error text shows up in several independent reports within a week, that page gets written before the trendier topic, because the people searching it are already blocked and already looking for an answer.

That is also why the site is small on purpose. Ten pages written from first-hand checks are worth more than a hundred rewritten release notes, and the second kind is what a search engine already has too much of. If you want to know what is coming, the queue is public on this page; if something is missing that you need today, say so and it moves up.

How this site is built

• Dated, not evergreen. Prices, limits and error messages change; each page carries the date its facts were last checked.
• No affiliate-driven rankings. We do not rank tools by who pays; if a link is ever monetised it will be labelled.
• Static and lightweight. No tracking scripts, no cookie banners, no client-side JavaScript.

Questions