Claude Code and Codex

Syncretic has no API key. It runs the command-line tool of your subscription, on this Mac, for each call.

Work with	Tool	Plans	Models offered
Claude	Claude Code	Claude Pro or Max	Opus, Fable, Sonnet, Haiku, by family (any version Claude Code reports resolves to its family); efforts Low, Medium, High, Extra high, Max
ChatGPT	Codex	ChatGPT Plus, Pro, Business or Enterprise	whatever Codex lists in its own model cache (refreshed whenever Codex runs), so new models appear without an app update; efforts as each model supports

Each provider keeps its own choice of models, so switching back finds them as they were. Defaults: Sonnet at medium effort for writing and building; Sonnet at low effort for routing; for ChatGPT, the first model Codex lists, and a “mini” one for routing when there is one.

The status row

Shown in Settings, in the search bar’s AI models tab, and in the Build panel when something’s wrong:

Claude Code isn’t installed. → Install opens a Terminal window and runs the official installer, so you can watch it.
Claude Code isn’t signed in. → Sign in runs the tool’s login and opens its page in your browser; the link is shown in case it doesn’t. Cancel stops it.
Too old for this model → Update: Claude Code updates itself (claude update, with its output shown as it goes); Codex is updated in Terminal with the command for the way it was installed.
Signed in: the account and version.

A failed call teaches the row: a sign-in error marks the tool signed out, an unknown model marks it outdated, and the fix is offered where the call failed (the trail, the Build turn) as well as here.

Effort

How hard the model thinks before answering. More effort is slower and more thorough; routing is a small decision, so it runs low by default. The Build panel’s model menu sets both for a conversation; the meta line under each answer says what was used.

Debugging

Debug builds accept SYNCRETIC_FAKE_AI=1 in the environment to exercise Ask and Build with canned answers and no tool at all.