JevCode — coasty-ai

A coding agent that runs in your terminal: you chat with it, and when you ask for a change it reads, edits and runs your code itself. It streams replies, works through tools such as searching, editing and running commands, and checks its own work with the fastest test or typecheck covering the change.

On the side. The README describes Jev as a small decision model that makes a few quick calls; on models where it helps, it spots a simple question so the reply comes back faster. It never blocks anything, and its key is optional — use /jev compact off style controls only exist elsewhere; here Jev is purely an accelerator.

npm i -g @coasty/jevcode to install, or npx @coasty/jevcode to try it without installing. Run jevcode in your project: the first time it asks for an OpenRouter key and saves it. You need Node 22.12 or newer on macOS, Linux or WSL 2; Homebrew, AUR, Nix and mise channels are documented in the install guide.

OpenRouter is the default (OPENROUTER_API_KEY), and --provider switches to openai, anthropic, xai, fireworks, meta or gemini, each with its own key. Export the key, put it in a .env file, or save it with jevcode login.

The README's recorded demo says two bugs were fixed in under ten seconds for about a tenth of a cent, but that is an illustration, not a price list. Real cost depends on your provider and model; the Jev key is optional, so the decision-model calls can be skipped entirely.

It acts without asking by default, and /undo reverts a step's changes to your files. Start it with jevcode --autonomy review to approve risky commands first, and add --agent-verify tests to have JevCode run your whole test suite as well as the fastest covering check.

It is MIT licensed and the README documents a demo project for a first run, a configuration reference and a CLI reference. The digest captured the whole README, but Jev's role is described only briefly there, so treat the decision-model assist as a small optimization rather than the reason to use the agent.