jevx

Find where Jev fits in your codebase — and change it.

Your code is full of hardcoded rules that are really judgment calls: regexes that guess what an error means, keyword lists that guess what a customer wants, slice(0, 3) that guesses what's best. jevx finds them and replaces them with a Jev decision (TypeSafe noul yes/no · choice · score) — keeping the old rule as the fallback.

npm i -g @vij-sameerb5/jevx          # or: npx @vij-sameerb5/jevx
cd your-project
jevx --dry-run         # see what it would change
jevx                   # change the strong fits, run your tests, undo anytime

60-second start

1. An AI key of your own — jevx never ships one:

   export XAI_API_KEY=…            # xAI (Grok)   — or OPENROUTER_API_KEY=…
   export TYPESAFE_API_KEY=…       # optional: Jev's own opinion on every scorecard

   Or put them in a file and point to it once: export JEVX_ENV_FILE=~/path/to/.env.
2. jevx --dry-run in your project → scorecards + red/green previews, nothing written.
3. jevx → strong fits are written, your tests run before and after, anything that breaks them is undone automatically. Review in VS Code / Cursor → Source Control. jevx undo restores all.

What happens

index your repo (local, free)
  → your AI reads the source, logic first, and names every spot that looks like a judgment
  → it re-checks each spot with the surrounding code and drops exact logic
  → scorecard: your AI + TypeSafe (Jev itself) + patterns learned from real Jev projects
  → writes fits ≥ 70% (keeps the old rule as fallback, updates callers, adds @typesafe-ai/sdk)
  → runs your tests / typecheck before and after; reverts what breaks

Commands

| | | | --- | --- | | jevx | find, judge, change strong fits, test | | jevx --dry-run | preview everything from 50% fit; write nothing | | jevx --min-fit 55 | also write fits from 55% (50–100; default 70). Under 70 warns, under 50 refused | | jevx --include-disagree | also write fits the sources disagree on (if they reach --min-fit) | | jevx undo | put back everything the last run changed | | jevx --fast | less AI reasoning while reading (cheaper, may miss spots) | | jevx --share | share anonymous outcomes to improve jevx (see Privacy) | | jevx --report-json out.json | save the full result locally | | jevx mcp install | use it from Claude Code, Claude Desktop, Cursor, … instead (their AI, no key needed) |

Claude Code, Claude Desktop, Cursor & more

jevx mcp install        # jevx mcp uninstall removes it again

Adds JevX to every supported app on this machine: Claude Code, Claude Desktop, Cursor, Windsurf, VS Code (Copilot), Gemini CLI and Codex (each config is backed up once; nothing else in it is touched). Restart the app and say "use jevx to find where Jev fits in this repo". The app's AI reads your code with jevx's tools, scores each spot and makes the changes — or, where it can't edit files (Claude Desktop), applies them with jevx_apply: backup, tests before/after, auto-revert. Say "use Jev where the fit is at least 55%" to go lower.

Claude Desktop, one click: double-click jevx.mcpb (or Settings → Extensions → Advanced settings → Install Extension), pick your project folder, and paste your TypeSafe key if you have one (Claude Desktop stores it as a secret). In chat, name other folders: "…in ~/code/my-app".

• The "not verified by Anthropic" warning is expected: JevX isn't in Anthropic's reviewed directory yet. The bundle is open source (this repo), MIT, contains no keys, and is built by pnpm build-mcpb with a content check — build it yourself if you prefer.
• Approvals: Claude Desktop asks per tool. Choose Always allow for the read-only tools (jevx_guide, jevx_scan, jevx_read, jevx_search, jevx_related) — they never write or send anything. Keep asking for jevx_apply / jevx_undo, which change files. jevx_read reads up to 20 files per call, so a scan needs few approvals.

Privacy

• Sent to your AI (xAI / OpenRouter, your key): your source files, secrets scrubbed. Never .env files, tests, builds or node_modules. You're asked once per provider.
• Stored by jevx (only with --share): one anonymous row per finding — the kind of code (e.g. "regex on an error message"), why Jev fits, the scores and what happened. Never code, file or function names, paths, string literals or your repo's name — scrubbed in code before sending. The dataset accepts inserts only; nobody can read it with the public key.
• .jevx/ in your project holds backups, the report and debug output. It ignores itself in git.

Cost

About 15–25 AI calls for a 50-file app (~$0.50–0.90 with Grok 4.6), on your key. The run stops at --budget tokens (default 400,000). Changed code calls Jev at runtime: set TYPESAFE_API_KEY in your app's environment.

Requirements

Node ≥ 20.10 · a TypeScript or JavaScript project · macOS, Linux or Windows.

MIT · changelog