jay-gentic v0.4.03.0: The Open-Source Agentic LLM Framework for Secure, Multi-Provider Deployments

Version 0.4.03.0 is now available. jay-gentic has evolved from a clever bash hack into a production-ready, modular agentic framework that runs on your infrastructure—no external API dependency required. With support for 8 new cloud providers, a 6-layer configuration precedence system, and hardened multi-provider compatibility, it’s the tool for organizations that need transparent, auditable AI assistance.

What Is jay-gentic?

jay-gentic is a modular, shell-native agentic LLM framework designed for security teams, DevOps engineers, and development organizations that need:

• Local execution — Run open-source models on your hardware via Ollama
• Multi-provider flexibility — Claude, OpenAI, Mistral, DeepSeek, Gemini, Perplexity, and more
• Transparent agent orchestration — Coordinate multiple specialized agents
• Auditable tool calling — Every command is logged and verifiable
• No external API lock-in — Works with cloud APIs or local models, your choice

At its core, jay-gentic is 21 composable bash modules + 3 Go binaries that decompose agent orchestration into testable, reusable pieces.

The Evolution: v0.4.0 → v0.4.03.0

v0.4.0 (April 4, 2026): The Modular Foundation

v0.4.0 introduced the architectural shift from a monolithic bash script to a composable framework:

Key systems:

• Agent system — Run, validate, and sandbox agent personas with YAML definitions
• Memory system — ACT-R activation scoring with episodic memory decay
• Permissions engine — Fail-closed enforcement with glob-style rules
• Tool execution — Input validation, injection prevention, 30s timeout
• Voice synthesis bridge — HTTP bridge to voice-server for audible responses
• Claude Code compatibility — 25+ bidirectional tool name mappings

v0.4.02.2 (April 22, 2026): Provider Stability

v0.4.02.2 hardened multi-provider compatibility across OpenAI, Gemini, and Ollama:

Fixes:

• ✅ Duplicate tool call message accumulation (OpenAI)
• ✅ Gemini authentication with provider-specific keys
• ✅ JSON validation and error handling for tool calls
• ✅ Streaming metadata parsing across all providers
• ✅ OpenAI whitespace trimming, duplicate tool registration prevention

Result: Reliable tool calling across all major providers, even with streaming enabled.

v0.4.03.0 (April 23, 2026): Multi-Provider Expansion & Config Architecture

v0.4.03.0 brings 8 new provider integrations, intelligent routing, and a 6-layer configuration precedence system:

8 New Providers:

Smart Routing for OpenCode Zen:

claude-*  → /v1/messages              (Anthropic format)
gpt-*     → /v1/chat/completions      (OpenAI compat)
*         → /v1/chat/completions      (Fallback)

This means you can route different model families to their native endpoints automatically—no manual configuration needed.

6-Layer Config Precedence:

Layer 1: Defaults (compiled-in)
Layer 2: ~/.claude/config.sh
Layer 3: ~/.jay-gentic/config.sh ← ENTERPRISE_MODE locks here
Layer 4: <project>/.claude/config.sh
Layer 5: <project>/.jay-gentic/config.sh
Layer 6: CLI arguments ← Always wins

This enables fine-grained control:

• Global defaults apply everywhere
• SMBs can lock organizational policies via ENTERPRISE_MODE
• Projects override globally, preserving global rules
• Individual commands override all

Multi-Layer Context Aggregation:

CLAUDE.md, AGENTS.md, and .jay-gentic.md files are aggregated across all directory levels, providing system-wide context without file duplication.

Core Features: A Technical Deep Dive

1. Agent Orchestration

Define agents in YAML and coordinate them:

# agents/security-analyst.yaml
name: security-analyst
model: sonnet
permissions:
  - "bash(grep*)"
  - "bash(find*)"
  - "jq(*)"
tools:
  - file-read
  - bash-execute

Then run them:

jay-gentic --agent security-analyst --task "Find all recent log anomalies in /var/log"

2. ACT-R Memory System

Memories score based on recency and frequency. Frequently accessed facts stay fresh; old facts decay. Backed by Go’s SQLite binaries for performance.

jay-gentic memory add "fact: AWS RDS uses IAM database auth"
jay-gentic memory add "fact: Always rotate credentials quarterly"

# Retrieve top facts by activation score
jay-gentic memory list --top 10 --decay-hours 24

3. Permission Enforcement

Glob-style rules block dangerous patterns before execution:

{
  "permissions": {
    "blocked": [
      "bash(rm -rf *)",
      "bash(*> /etc/*)",
      "bash(curl | bash)"
    ]
  }
}

4. Multi-Provider Execution

Run the same prompt against different providers for comparison:

# Run against local model
jay-gentic --server ollama --model llama3.2 --prompt "Explain Zero Trust"

# Run against Claude
jay-gentic --server claude --model claude-opus-4-6 --prompt "Explain Zero Trust"

# Run against OpenCode Zen
jay-gentic --server opencode --model claude-sonnet-4-5 --prompt "Explain Zero Trust"

5. Comprehensive Logging

All operations logged as structured JSONL:

# View logs
tail -f ~/.jay-gentic/logs/jay-gentic.jsonl

# Query logs
jay-gentic log query "error" --limit 100

6. Tool Execution with Input Validation

Prevents command injection and path traversal:

# This works:
echo "What files exist in /tmp?" | jay-gentic --one-shot "List files"

# This is blocked:
echo "/tmp; rm -rf /" | jay-gentic --one-shot "Execute this"

7. Rate Limiting

Prevent runaway tool loops:

jay-gentic ratelimit check --max-requests 100 --window 3600

8. Voice Synthesis Bridge

Audible responses via HTTP bridge to voice-server:

jay-gentic --voice --voice-id "en-US-Neural2-A" "Explain the Zero Trust model"

Security for SMBs: Why This Matters

For small and medium-sized organizations, jay-gentic solves critical problems:

1. No API Dependency

Your security analysis stays on-premises. No cloud logging, no vendor lock-in. Run Llama 3.2 locally for free.

2. Auditable Every Step

Every prompt, every tool call, every model invocation is logged. Full compliance trail for regulatory requirements (SOC 2, HIPAA, GDPR).

3. Cost-Effective at Scale

• Local models (Ollama): $0 per inference
• Cloud APIs (Claude, OpenAI): Pay per token
• Mix and match: Use free local models for routine tasks, reserve paid APIs for complex reasoning

4. Transparent Orchestration

Unlike black-box AI assistants, you see exactly:

• Which agent is running
• Which tool is being called
• What permissions were checked
• Why a request was blocked

5. Multi-Provider Flexibility

No vendor lock-in. Plug in your preferred LLM:

• Local: Llama, Mistral, DeepSeek via Ollama
• Cloud: Claude, GPT-4o, Gemini, Perplexity
• Routing: Different models for different tasks

Getting Started: Installation & First Run

Prerequisites

# Required
bash --version     # 4.0+ (macOS: brew install bash)
curl --version     # Any recent version
jq --version       # 1.6+

# For Go helpers (build once)
go version         # 1.21+

# For local LLM (optional but recommended)
ollama --version

Step 1: Clone jay-gentic

cd ~/Projects
git clone https://github.com/delphijc/jay-gentic.git
cd jay-gentic
chmod +x jay-gentic

Step 2: Build Go Helper Binaries

# Build memory, settings, sanitize helpers
cd helpers/jg-memory && go build -o ../bin/jg-memory . && cd ../..
cd helpers/jg-settings && go build -o ../bin/jg-settings . && cd ../..
cd helpers/jg-sanitize && go build -o ../bin/jg-sanitize . && cd ../..

# Verify
ls -la helpers/bin/

Step 3: Pull Local Models (Optional)

# Install Ollama first: https://ollama.ai

# Pull lightweight models
ollama pull llama3.2           # ~2GB, good for general tasks
ollama pull llama3.2:3b        # ~2GB, lightweight
ollama pull mistral            # ~5GB, strong reasoning

Step 4: Initialize Data Directory

# Create ~/.jay-gentic/ with standard directories
./jay-gentic --version    # Triggers auto-init

# Verify
ls -la ~/.jay-gentic/
# Should contain: memory/ tools/ hooks/ agents/ skills/ config/ logs/

Step 5: Test Local Execution

# Query using Ollama
./jay-gentic --server ollama --model llama3.2 --one-shot "What is Zero Trust security?"

# One-shot with Unix pipe
echo "List common cybersecurity frameworks" | ./jay-gentic --server ollama --model llama3.2 --one-shot "Answer:"

Step 6: Configure a Cloud Provider (Optional)

# Set your API key
export ANTHROPIC_API_KEY="sk-ant-..."

# Test Claude
./jay-gentic --server claude --model claude-opus-4-6 --one-shot "Hello, what is your version?"

Step 7: Run the Test Suite

# Install bats
brew install bats-core  # macOS
# or: sudo apt install bats  # Linux

# Run tests
cd ~/Projects/jay-gentic
bats tests/phase-1/*.bats
bats tests/security/*.bats

# Run everything
bats tests/**/*.bats

Architecture at a Glance

jay-gentic (main entry)
├── jay-gentic.d/
│   ├── systems/      (subsystems: agents, memory, hooks, tools)
│   ├── tools/        (command implementations)
│   └── defaults.sh   (built-in config)
├── lib/              (21 composable bash modules)
│   ├── jg-agent.sh
│   ├── jg-memory.sh
│   ├── jg-permissions.sh
│   ├── jg-execute.sh
│   ├── jg-log.sh
│   └── ... (15 more)
├── helpers/bin/
│   ├── jg-memory     (SQLite ACT-R episodic memory)
│   ├── jg-settings   (fast JSON config parsing)
│   └── jg-sanitize   (text escape/injection filtering)
├── config/
│   └── model-routes.json  (provider routing table)
├── tests/            (33 BATS test files)
└── docs/wiki/        (12 documentation pages)

Use Cases: What You Can Do Now

1. Security Incident Triage

# Analyze a suspicious log file
cat /var/log/auth.log | jay-gentic --agent security-analyst --prompt "Identify suspicious login attempts"

2. Code Review with Local Models

# Review a patch without uploading to cloud
git diff HEAD~1 | jay-gentic --server ollama --model mistral --prompt "Review this code for security issues"

3. Multi-Provider Comparison

# Compare responses across models
for model in "llama3.2" "mistral"; do
  echo "=== $model ===" 
  echo "Explain IAM roles" | jay-gentic --server ollama --model $model
done

4. Automated Compliance Checks

# Run with custom agent & logging
./jay-gentic --agent compliance-checker \
  --permissions /etc/jay-gentic/policies.json \
  --log-file /var/log/compliance-checks.jsonl

5. DevSecOps Integration

# Run in CI/CD pipeline
jay-gentic --server opencode --model big-pickle \
  --no-stream \
  --output json \
  < security-requirements.txt

Roadmap: What’s Coming

The team is actively working on:

• Plugin marketplace — Community-contributed agents and skills
• Workflow engine v2 — DAG-based execution with conditional branching
• Memory federation — Shared memory across jay-gentic and Sam’s SQLite store
• GPU-aware routing — Automatic model selection based on available hardware
• Web dashboard — Monitoring and observability UI

Try It Now

git clone https://github.com/delphijc/jay-gentic.git
cd jay-gentic
chmod +x jay-gentic

# Run with local model (free)
./jay-gentic --server ollama --model llama3.2

# Or with Claude
export ANTHROPIC_API_KEY="sk-ant-..."
./jay-gentic --server claude --model claude-opus-4-6

No Python. No Node. No Docker. Just bash, curl, jq, and an LLM.

Learn More

• GitHub Repository — Code, issues, contributions
• Provider Setup Guide — Configure cloud providers
• Architecture Deep Dive — Module organization and design
• Security Testing — See the threat model and test coverage

jay-gentic is MIT licensed. Contributions, issues, and feedback are welcome.

For SMBs building secure, auditable AI systems without cloud lock-in, jay-gentic is built for you.