> ## Documentation Index > Fetch the complete documentation index at: https://reikon.dev/docs/llms.txt > Use this file to discover all available pages before exploring further. # Why Reikon > Take the helm of your codebase — the seat where you decide what's worth doing and judge whether your agents' work actually did it. # Why Reikon **Take the helm of your codebase.** Your coding agents can already write code faster than you can decide what's worth doing — and judge whether what comes back actually did it. That judgment, not the typing, is the bottleneck now. Reikon is the seat you make it from: where you decide what to steer toward, and judge every change your agents hand back. It's local-first and a product, not a platform: download it, open a folder, get a health score in seconds. No account, no seats, no sales call. Nothing leaves your machine unless you export a report or point it at an AI tool you already run. ## The wheel A decision has a lifecycle, and Reikon's spine *is* that lifecycle, as a wheel: * **Goals — set intent.** Tell Reikon what you're steering toward ("fewer `useEffect`s", "no file over 500 lines", "keep coverage above 80%") in plain language. It tracks each goal deterministically, forever after. * **Review — judge the act.** Your agent flags risky hunks and comments on lines over MCP; you reply, resolve, and **dispatch** fixes back. The diff returns for you to accept or send back. * **Verify — judge the result.** Is it actually true now? Reikon reads your coverage report and surfaces the riskiest files with no tests — a "test these first" list you can hand straight to an agent. * **Architecture — judge the form.** Is the structure worth keeping? Layer-violation findings — the leg that closes the loop back into new Goals. You stay at the helm the whole way around: Reikon surfaces what's wrong and judges what comes back; the agent does the rowing. ## The loop — and dispatch Reikon ships an MCP server with **write** tools, not just read ones — that's the difference. Your agent (Claude Code, Cursor, Codex, anything that speaks MCP) flags a hunk, comments on your notes, and — when you ask — **fixes** what you dispatched: 1. You leave a note, or your agent flags one. You hit **Copy prompt** — that one act marks the note fix-requested *and* gives you a kickoff prompt. Paste it into your agent, or tell an MCP-connected agent to *"work Reikon's dispatch queue"* and it pulls the queue itself (see the [`reikon-queue` skill](../guides/mcp-claude-code)). 2. The agent fixes it **in its own environment** and leaves the change in your working tree. 3. The diff returns to **Review**, linked to the flag. You **accept** it, or **send it back** with a reason that rides along on the re-dispatch. Reikon emits the instruction and reviews the result — it **never applies a fix itself.** It's a cockpit, not an editor. And the trust runs one way: a returned fix only counts once there's a real change in the working tree, never on the agent's word for it. The same review notes live in the desktop app, the `rei` CLI, and your agent — in sync, on your machine. See the [MCP guides](../guides/mcp-claude-code) to wire it up. ## Why this gets stronger as agents improve Most tools in this space bet on the *agent* being the product. Reikon bets on the *judgment directing it*. The better your agents get at writing code, the more the bottleneck becomes deciding what's worth doing and checking what they hand back — which is exactly the seat Reikon builds. Insight originates the work, a human reviews every returning diff, nothing leaves your machine, and Reikon never executes. No hosted PR-bot can hold that cell. ## What it measures — and shows its math Underneath the loop is deep, deterministic git analysis. Reikon computes a health score from six signals (bus factor, commit quality, TODO density, duplicate code, churn stability, dependency health), estimates tech debt per file from complexity × churn × size, flags files only one person understands, and — unusually for this category — shows the actual math. See [Formulas & Methodology](../reference/formulas) to verify any number yourself instead of taking it on faith. ## What Reikon won't do The boundaries are the point — they're what keep it private and honest: * **No hosted service, no account, no PR bot.** Branch and diff review run on local git only — no GitHub/GitLab API, no tokens, nothing to authorize. * **No semantic search or vector index.** Every signal is deterministic and reproducible, not an embedding you have to trust. * **No AI-written source from Reikon.** When you dispatch, *your* agent makes the change in its own environment and you review the diff. Reikon emits and judges; it never writes your code. Its own write tools touch review notes only. ## AI is optional, and keyless Reikon's AI features (health explanations, goal drafting, summaries) work by **driving a coding-agent CLI you already have** — Claude Code, Codex, or Gemini, whichever you've already signed into. Reikon never holds a key or makes an AI network call; the CLI owns its own auth. Skip AI entirely and every core feature still works — the health score, the loop, the analysis, the CLI. AI is additive, never load-bearing.