Skip to main content

​
AI Features

Fully useful without AI. Every core feature — health score, the review loop, code quality, dependency audits, the CLI, MCP integration — works with zero AI setup. AI adds explain / summarize / draft capabilities on top; it never gates anything load-bearing.

​
How it works — keyless, driven

Reikon never holds an API key or makes an AI network call. Instead it drives a coding-agent CLI you already have. You pick which one in Settings → AI, where Reikon scans for installed tools:
  • Claude Code, Codex, or Gemini — a coding-agent CLI you’ve already signed into. Reikon spawns it for a one-shot completion; the tool owns its own auth, billing, and model (Reikon just shows you which model answered).
Your prompts go to whichever CLI you picked, under that tool’s own privacy terms — Reikon is never in the path. You can also turn AI off entirely (None); every core feature still works.
The one-shot invocation flags per CLI are best-effort and can drift as each tool changes its headless interface — if a feature stops answering, check that the CLI runs from your shell.

​
What AI unlocks

  • Duplicate contributor detection — suggests merges when the same person appears under multiple git identities (different name spelling, work vs. personal email)
  • Health signal explanations — plain-English reasoning behind why a signal scored the way it did
  • Commit message suggestions — rewrites for messages flagged as low-quality
  • TODO categorization — sorts TODOs into bug / debt / feature / question
  • Duplicate code explanations — what a flagged duplicate block actually does
  • Circular dependency explanations — why a detected import cycle matters
  • Executive summary / standup copy — a 2–3 sentence plain-English summary of current repo state, written for sharing
  • Snapshot narrative — what changed between two points in time, in prose
  • Commit convention advice — suggestions for adopting more consistent commit message conventions
  • Goal drafting — turns a plain-language goal (“fewer useEffects”) into a precise, countable rule you confirm before saving

​
Privacy model

Nothing reaches an AI provider unless you’ve picked a CLI yourself and the feature is actually invoked — and then it goes to that CLI’s provider under its own terms, never through Reikon. Turn the include-file-content toggle off in Settings → AI to send metadata only, no code. Reikon caches AI responses per-project so re-asking the same question (on unchanged data) doesn’t repeat the same call; clear the cache any time from the same panel. Every AI explain popover also has a Regenerate button next to Copy that re-asks the model with the cache bypassed for that one explanation, when you want a fresh take rather than the stored answer.

​
Which model answers

Reikon runs whatever model your CLI is configured to use — there’s no model picker and no download step. The model that answered is read back from the CLI’s response and shown in Settings → AI after the first AI call (Claude Code reports it directly; others fall back to “uses your tool’s configured model”).