Developers

The biggest cohort in the industry right now: developers who live in a terminal but haven’t gone deep on the agentic meta. This slot optimises for them - maximum capability, minimum configuration.

Why Claude Code takes gold

Claude Code pips Codex here for a better out-of-the-box experience. Any user can pick it up and get roughly 80% of the experience an obsessive with bespoke AGENTS.md files, skills and hooks would assemble by hand: subagents that run in the background by default, background agents that finish a task by committing and opening a draft PR, sandboxing on every platform, and the best multi-agent orchestration paradigm currently shipping. The users agree - it’s the satisfaction leader by a landslide (46% most-loved, 91% CSAT, the highest of any AI coding tool) - and the defaults are genuinely tuned: Anthropic deleted the large majority of its system prompt this summer with no measurable eval loss, which is more than most harnesses can say about their accumulated cruft.

Two concessions, printed rather than buried. Codex CLI beats it on terminal benchmarks per token - roughly 4× more token-efficient for equivalent tasks in one comparison - so if you’re paying per token rather than riding a subscription, the maths may point the other way. And the trust ledger has an entry: Claude Code spent months covertly fingerprinting requests via invisible Unicode characters, removed with no changelog entry. The community largely shrugged (‘wait until you hear about web browsers’). We noted it anyway. It’s strictly terminal-only in practice, too - the desktop app is not worth your disk space, which is exactly why this tool is gold here and bronze in Most people.

The runners-up

Codex is the open-source, more-customisable, nearly-as-capable alternative and the better long-term bet if you expect to outgrow defaults - plus the benchmark leader for terminal work. Sandbox it: the family’s destructive streak is documented, and full-access mode has eaten a home directory.

Cursor rounds out the podium for developers who want the same power inside an editor rather than a terminal - and its tab completion remains the best ambient AI experience in any IDE.

Connoisseur’s note: OpenCode has the highest engagement-per-thread of any harness in our community research, a system prompt a fraction the size of anyone else’s, and agent definitions that are enforced rather than advisory. Small ecosystem, real ideas. One to watch.

What to do now

If you’re terminal-native and just want the strongest defaults today: Claude Code, and learn its skills and hooks primitives before building your own - most of what you’d build is already in the box. Read the primitives guides first. Your config file will thank you.