jev-x-kit

Universal TypeSafe Jev / OpenJev autonomous decision, deep research, ultra-planning & self-improving agent framework — packaged as a Claude Code plugin/skill, an MCP server, and a standalone CLI.

Built for Claude Code, Cursor, Codex, OpenCode, Continue.dev, Ollama / vLLM, and every MCP-compatible agent. Runs 100% free and offline: no API key, no GPU, no cloud required.

Install as a Claude Code plugin (30 seconds)

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

Then point Claude Code at this folder as a plugin (.claude-plugin/plugin.json is already wired up: it registers the jev skill and the jev-super-agent MCP server with 21 tools). See skills/jev/SKILL.md for the command cheat-sheet Claude reads to decide when to reach for Jev instead of free-text reasoning.

Why this exists (6 root problems)

1. Latency wall — LLMs take 3–30s for a simple yes/no routing decision.
2. Token cost explosion — agent loops burn $10–$50 per session on micro-decisions.
3. JSON schema breakage — free-text models hallucinate parameters and break parsers.
4. Context rot — "summarized" logs hallucinate file paths, commands and error codes.
5. Single-pass blind spot — a model can't adversarially review its own plan.
6. Vendor lock-in — paid APIs only, no offline path.

What this kit does

It inserts a System 1 decision layer in front of every micro-decision:

• Non-autoregressive primitives: Choice (up to 255 options), Score (fractional 1–10), Noul (calibrated probability in [0, 1]) — never free text, schema-validated, zero output-token cost.
• "BELKİ" Gatekeeper: confidence > 0.85 executes directly at $0 LLM cost; 0.60–0.85 splits into speculative sub-decisions; below 0.60 escalates to System 2 or a human.
• Speculative fan-out: N independent questions are evaluated in one parallel pass — extra questions add no latency and no cost.
• Winnow lossless compaction: irrelevant log lines are deleted, never summarized. File paths, commands, error codes, URLs and diff headers are always preserved byte-identically.
• Adversarial red-teaming: a hostile reviewer generates concrete anti-theses; the Jev loop arbitrates the safest route with residual risks.
• RLVR self-improvement: tsc + npm test pass/fail records ±1 rewards into .jev-skill-memory.json; verified wins auto-tune the gatekeeper thresholds.

Quickstart (zero config, offline)

npm install
npm run build
npm test            # 27 unit tests (core 20 + datacenter 7) on the deterministic offline simulator
npm run test:hub    # 7 datacenter tests (source registry, robots, license engine, FTS store, query tiers)
npm run smoke       # full end-to-end MCP smoke test (50 checks)

Everything works immediately with no network, no GPU and no API key — the deterministic offline simulator keeps every primitive valid and schema-safe.

MCP server setup

Add this to your agent's MCP config (mcp-config.json, claude_desktop_config.json, opencode.json, Continue.dev, Codex):

{
  "mcpServers": {
    "jev-super-agent": {
      "command": "node",
      "args": ["C:\\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"
      }
    }
  }
}

Free local backends (auto-detected in order)

Backend	Command	Cost
OpenJev on vLLM (razorback16/openjev)	docker run --gpus all -p 8000:8000 razorback16/openjev	$0
LayA (NandhaKishorM/laya)	python -m laya.serve --port 8000	$0
Ollama (System-2 LLM for planner/red-team)	ollama serve (default :11434)	$0
Vercel AI Gateway free tier	set VERCEL_AI_GATEWAY_KEY	free tier
TypeSafe Jev native	TYPESAFE_JEV_API_KEY + TYPESAFE_JEV_NATIVE=1 (required — see below)	$0.042 / 1M in, $0 output

Set JEV_BACKEND_PROVIDER to auto (default), typesafe_jev, openjev_local, laya_local or heuristic (fully offline).

TYPESAFE_JEV_NATIVE=1 is not optional for the hosted API. TypeSafe's real API is a single POST /v1/systemone endpoint (state + named questions in, named answers + token usage out) — there is no /chat/completions endpoint at api.typesafe.ai, so without the native flag the chat-proxy transport 404s and every call silently falls back to the offline heuristic. Verified live on 2026-09-21: a 12-candidate jev_research run cost $0.00016874 in real TypeSafe API usage. Note the hosted API only answers Choice/Score/Noul — it has no chat/completion endpoint, so point JEV_LLM_BASE_URL at Ollama/vLLM/etc. regardless, for planner/red-team/research-synthesis text generation.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗