JevXagent — AI skill for Claude Code

jevXagent — When JEV Meets LLM Agent

**A local LLM proxy / AI-agent middleware that sits between an LLM agent (e.g. Claude Code) and its Anthropic-compatible model API.**

The primary model (Claude) keeps the **global context** and does the **complex reasoning**. A fast auxiliary model (**JEV**) is used as a **decision coprocessor** for suitable low-complexity decisions. Built-in **decision routing**, **telemetry**, a **statistics dashboard**, and **benchmarking** let you measure whether routing actually reduces latency, token usage, and cost for agentic workflows.


**Status:** Phase 1 (transparent proxy) works and is verified against a real Claude Code session. Routing, interception, verification and the MCP tool are implemented. **JEV routing is OFF by default** until a real JEV API key is configured. Every "savings" number is an **estimate** and clearly labelled — no measured number is ever invented.

---

Table of contents

1. What is jevXagent?
2. The main goal
3. What it can do (capability audit)
4. How it works
5. Quick start: from clone to Claude Code
6. Using it with Claude Code
7. Two modes: relay-only vs routing
8. Configuration reference
9. Decision routing API
10. Statistics dashboard
11. Benchmark mode
12. MCP tool: `jevx_decide`
13. Testing
14. Security & privacy
15. Project structure
16. Roadmap
17. Troubleshooting
18. License

---

What is jevXagent?

An agentic workflow (coding, research, automation) is not one giant reasoning task. It is a long sequence of **many small decisions**:

• What is this object? Which category / option / tool / argument?
• Is this result relevant? D