jevXagent v0.2

Cut AI Agent Response Time by 70% & Save 80% on Tokens

The transparent proxy that makes your AI coding agents faster, cheaper, and smarter

Quick Start • Features • How It Works • Stats • Docs

---

What is jevXagent?

jevXagent is a transparent proxy that supercharges your AI coding agents (Claude Code, Codex, Kilo, Cline) by routing prompts through Jev's blazing-fast decision engine before the main LLM generates a response.

The Problem

Your AI agent wastes time and tokens doing the same classification work over and over:

• "Is this a DevOps question or a bug fix?"
• "Should I search the codebase or generate code?"
• "What's the user actually asking for?"

Every prompt burns 200-300 expensive output tokens just thinking about what to do.

The Solution: Line J

jevXagent implements "Line J" - injecting Jev's 120ms decisions directly into the agent's context, so it sees what was already decided before generating:

Without jevXagent: 2,800ms, 250 tokens
User → Agent → [classify + think + generate] → Response

With jevXagent: 950ms, 50 tokens  
User → Jev (120ms) → Line J → Agent → [generate only] → Response

Result: 3x faster, 80% fewer output tokens, same quality.

---

Real Performance

Live Statistics from Production Use

Updated weekly - see statistics/ folder

Metric	Without jevXagent	With jevXagent	Improvement
Response Time	2,400ms	890ms	2.7x faster
Output Tokens	240 tokens	48 tokens	80% reduction
Cost per Request	$0.0072	$0.0018	75% cheaper
Time to First Token	850ms	310ms	63% faster

---

Features

Core Capabilities

• Zero-Config Proxy - One command to start, works with all your agents
• 4 Agent Support - Claude Code, Codex, Kilo Code, Cline (auto-detected)
• Line J Injection - Jev decisions as system context (the secret sauce)
• Fail-Open Design - Jev timeout? Agent still works perfectly
• Real-Time Metrics - See latency and token savings per request

Monitoring & Analytics

• Live Console Display - Track every request's performance
• JSONL Statistics - Machine-readable logs for analysis
• Ctrl+S Snapshots - Instant summaries of cumulative stats
• Weekly Reports - Automated statistics visualization

Production-Ready

• Streaming Support - SSE with token extraction
• Multi-Turn Aware - Conversation context preserved
• Media Bypass - Images/PDFs skip Jev automatically
• Secure by Default - No credential logging, localhost only

---

Quick Start

Installation

# Clone and install
git clone https://github.com/j1s4nn/jevXagent.git
cd jevXagent
pip install -e .

First Run (Interactive Setup)

python -m jevxagent

Answer a few prompts:

1. Select your installed agent (Claude Code, Codex, etc.)
2. Enter Jev API key
3. Configure stats path
4. Done! Config saved to ~/.jevxagent/config.json

Daily Usage

Terminal 1 - Start the proxy:

python -m jevxagent

You'll see:

==================================================
jevXagent v0.2
==================================================
Mode: Project
Agent: claude-code
Jev: ON
Proxy: http://127.0.0.1:9099

Ctrl+S  Save statistics snapshot
Ctrl+C  Stop jevXagent
==================================================

Proxy running. Waiting for requests...

Terminal 2 - Use your agent normally:

# For Claude Code / Kilo
export ANTHROPIC_BASE_URL=http://127.0.0.1:9099
claude

# For Codex  
export OPENAI_BASE_URL=http://127.0.0.1:9099
codex

That's it! Your agent now routes through jevXagent automatically.

---

How It Works

The Line J Architecture

Traditional Agent (Slow):

User: "restart the pod"
  ↓
Agent LLM: "Let me think... this is DevOps... kubernetes... 
            I should generate a kubectl command... checking best 
            practices... here's the command with explanation..."
  ↓ 2,400ms, 240 tokens
Response: [Long explanation + code]

With jevXagent (Fast):

User: "restart the pod"
  ↓
Jev: Classifies → "DevOps, kubectl, urgent" [120ms]
  ↓
Line J: Injects context into agent request
  ↓
Agent LLM: *Sees decisions already made*
           "kubectl rollout restart..."
  ↓ 890ms, 48 tokens  
Response: [Concise code only]

What is "Line J"?

Line J is the architectural innovation that makes this possible. Instead of making Jev's decisions available to the agent, we inject them directly into the agent's system context:

{
  "messages": [
    {
      "role": "system",
      "content": "SYSTEM EXECUTION CONTEXT:\nCategory: DevOps\nUrgency: High\nAction: Generate kubectl script only"
    },
    {
      "role": "user", 
      "content": "restart the pod"
    }
  ]
}

Now the agent sees what Jev decided before generating, eliminating redundant classification and reasoning.

---

Live Statistics

Real-Time Console

Every request shows detailed metrics:

──────────────────────────────────────────────────
Request #42 | Trace: a3f8b2c1
Jev: success | 142ms
  Tokens: in=31 out=0
Agent: 1120ms
  TTFT: 310ms
  Tokens: in=1842 out=126 total=1968
Total: 1289ms
──────────────────────────────────────────────────

Ctrl+S Snapshots

Press Ctrl+S anytime for cumulative statistics:

==================================================
STATISTICS SNAPSHOT
==================================================
Total Requests: 347
Jev Enabled: 347
Bypassed: 23
Avg Jev Latency: 128ms
Avg Total Latency: 1,056ms
Saved to: ./jevxagent_stats.jsonl
==================================================

Weekly Reports

Generate visual statistics reports saved to statistics/:

# Generate weekly report
python scripts/generate_stats_chart.py

See statistics/README.md for details.

---

Use Cases

Perfect for:

• High-Volume Development - Save 80% on token costs across your team
• Performance-Critical Workflows - Cut response times from 3s to 1s
• Token Budget Management - Track and optimize AI spending
• Production Monitoring - Real-time visibility into agent performance

Works Best With:

• Coding tasks with clear classification (bugs, features, refactors)
• DevOps automation prompts
• Repetitive agent workflows
• Multi-agent orchestration systems

---

Advanced Configuration

Jev ON/OFF Toggle

Compare performance by toggling Jev:

# Edit ~/.jevxagent/config.json
{
  "jev_enabled": false  # Set to false to disable
}

Restart the proxy. Now you can measure the exact improvement Jev provides.

Custom Stats Path

Per-project statistics:

cd my-project
python -m jevxagent
# Stats saved to: my-project/jevxagent_stats.jsonl

Environment Variables

# Override config location
export JEVXAGENT_CONFIG=/path/to/config.json

# Change proxy port
export JEVXAGENT_PORT=9100

---

Documentation

• Quick Reference - Cheat sheet for daily use
• Usage Guide - Comprehensive walkthrough
• Implementation Report - Technical deep dive
• Architecture Diagrams - Visual architecture reference

---

Testing

Run the test suite:

pytest tests/ -v

Result: 19/19 tests passing

Coverage includes:

• Agent adapter detection
• Line J context injection
• End-to-end request flows
• Fail-open error handling
• Statistics persistence
• Multi-turn conversations

---

Roadmap

• Core proxy architecture
• 4 agent adapters
• Line J context injection
• Real-time metrics
• Statistics persistence
• Automated weekly reports
• Web dashboard (live metrics)
• Model-specific pricing
• Docker containerization
• Multi-agent switching
• Prometheus metrics export

---

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

Development Setup

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run linting
black src/ tests/

---

License

MIT License - see LICENSE for details.

---

Acknowledgments

• Jev by TypeSafe.ai - The fast decision engine powering jevXagent
• Anthropic, OpenAI - For the amazing agent APIs
• The open-source community - For the incredible tools we build upon

---

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Email: myprojectjisan@gmail.com

---

Speed up your AI agents today

Get Started • View Stats • Read Docs

Made with care by j1s4nn

Star us on GitHub if jevXagent helps you!