Your agent doesn't need
Opus to say yes.

~19x Fewer Tokens.
~11x Fewer Tool Calls.

*Measured on 3 real runs against live free APIs — not marketing copy. See the full methodology and reproduce every number yourself.

We took the boring
path on purpose.

Install as a Claude Code plugin

Thirty seconds. No account, no key required.

git clone https://github.com/Kadihx/jev-x-kit.git
cd jev-x-kit && npm install && npm run build

npm test            # 43 unit tests, offline simulator
npm run smoke       # 69-check end-to-end MCP client test

.claude-plugin/plugin.json is already wired up — it registers the jev skill and the jev-super-agent MCP server with 33 tools. Point Claude Code, Cursor, OpenCode or Continue.dev at this folder as a plugin, or hand-wire the MCP config yourself:

{
  "mcpServers": {
    "jev-super-agent": {
      "command": "node",
      "args": ["/path/to/jev-x-kit/dist/index.js"],
      "env": {
        "JEV_BACKEND_PROVIDER": "auto",
        "OPENJEV_BASE_URL": "http://localhost:8000/v1",
        "JEV_LLM_BASE_URL": "http://localhost:11434/v1"
      }
    }
  }
}

jev_research vs. vanilla Claude Code

One jev_research call fans four channels out outside Claude's context, ranks them with a $0 local Noul primitive, and returns only the ranked shortlist plus a synthesized brief.

Every number is either measured on real live free APIs, or derived from this repo's own crawled corpus — nowhere is a number invented. Wall-clock speed-up assumes 3s/manual turn (conservative, change it and re-run if you disagree). Reproduce any row yourself with node scripts/benchmark-vs-vanilla.mjs "your query". Full methodology in BENCHMARK.md.

batch() vs. one-at-a-time: 4.24x

Same 10-question battery, same real typesafe_jev backend, two ways of calling it: naive sequential calls (what a raw API integration looks like without the kit) vs. jev-x-kit's own backend.batch() (same-state questions merge into one HTTP call, different-state questions run concurrently).

LayA turns out to really exist — an open-source, self-hosted 421M ModernBERT + RLCD decision model (github.com/NandhaKishorM/laya) — so we installed it and ran the exact same 10-question battery against it, on this (GPU-less, CPU-only) machine:

Not a clean win for either side: this machine has no GPU (LayA publishes ~33–38ms/question on GPU, so these are CPU-bound numbers), and the battery is open-domain trivia — LayA's own model card lists its strengths as email triage, moderation and classification, not general knowledge, so this plays to a hosted LLM API's strengths more than LayA's. Reproduce it yourself with scripts/laya-benchmark.py; full numbers in artifacts/laya-benchmark-report.md.

Every number here is a real local measurement against the live TypeSafe Jev API, not a projection. Reproduce it yourself with node scripts/backend-benchmark.mjs (needs TYPESAFE_JEV_API_KEY in .env for the typesafe_jev row). Full methodology in artifacts/backend-benchmark-report.md.