The Local-First AI Site Reliability Engineer for WordPress
A privacy-first AI assistant that runs entirely in your browser, helping you diagnose and fix WordPress issues through natural language commands.
WP-Agentic-Admin transforms your WordPress admin panel into an intelligent command center. Instead of navigating through multiple screens to diagnose issues, you simply describe your problem in plain English:
"My site is throwing 500 errors"
The AI assistant will:
- Read your error logs
- Identify the problematic plugin or theme
- Propose a fix
- Execute it with your approval
All of this happens locally in your browser - no data is sent to third-party AI services, no API costs, complete privacy.
- 100% Local AI: Uses WebLLM to run Qwen 3 1.7B (default) or Qwen 2.5 7B directly in your browser via WebGPU
- Privacy-First: No admin data ever leaves your device - GDPR compliant by design
- Zero Server Costs: No GPU infrastructure needed - computation happens on the client
- Persistent AI: Model stays loaded across page navigations using Service Worker technology
- WordPress Abilities API: Natively integrates with WordPress's official Abilities API
- Extensible: Third-party plugins can register custom abilities
- Natural Language Interface: Describe problems in plain English, get intelligent solutions
- ReAct Loop: LLM decides which tools to use based on observations, adapting in real-time
WP-Agentic-Admin uses a ReAct (Reasoning + Acting) pattern where the AI decides tool selection one action at a time:
┌─────────────────────────────────────────────────────────────┐
│ Your Browser │
│ ┌───────────────────────────────────────────────────────┐ │
│ │ Local AI (WebLLM) │ │
│ │ Qwen 3 1.7B (default) / Qwen 2.5 7B │ │
│ │ │ │
│ │ ReAct Loop: LLM decides tools based on observations │ │
│ └───────────────────────────────────────────────────────┘ │
│ │ │
│ Tool Calls │
│ ▼ │
│ ┌───────────────────────────────────────────────────────┐ │
│ │ WordPress Abilities API │ │
│ │ Standardized capability discovery & execution │ │
│ └───────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────┘
│
REST API
▼
┌─────────────────────────────────────────────────────────────┐
│ WordPress Server │
│ ┌───────────────────────────────────────────────────────┐ │
│ │ Registered SRE Abilities (14 total) │ │
│ │ - Error log reading - Plugin management │ │
│ │ - Cache management - Site health diagnostics │ │
│ │ - Database optimization - Cron/rewrite/transient mgmt │ │
│ └───────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────┘
User: "My site is slow"
↓
AI: "I should check site health first"
↓
Tool: site-health → Returns database is 2.5GB
↓
AI: "Database is large, I should optimize it"
↓
Tool: db-optimize → Optimizes 15 tables
↓
AI: "Your database was causing slowness. I optimized it and saved 125MB."
The AI decides one action at a time, observing results and adapting its strategy.
Messages are routed through a simple 2-tier system:
-
Workflow detection — keyword matching triggers pre-defined multi-step workflows:
- "full cleanup" / "site cleanup" / "maintenance" → cache-flush → db-optimize (conditional) → site-health
- "performance check" / "health check" / "check site" → site-health → error-log-read (conditional)
- "audit plugins" / "check my plugins" → plugin-list → site-health
- "database maintenance" / "optimize database" → db-optimize → cache-flush
-
ReAct loop — everything else (questions, actions, diagnostics) goes through the AI-driven adaptive loop.
- WordPress 6.9+ (includes the Abilities API)
- PHP 8.2+ (8.3 recommended)
- Modern browser with WebGPU support (Chrome 113+, Edge 113+)
- Pretty permalinks enabled (required for REST API)
Click the badge above to launch a fully working WordPress instance with WP Agentic Admin pre-installed — no setup required. Requires a browser with WebGPU support (Chrome or Edge).
- Download and install WP-Agentic-Admin
- Navigate to "Agentic Admin" in your WordPress admin menu
- Wait for the AI model to download (one-time, ~1.2GB for Qwen 3 1.7B or ~4.5GB for Qwen 2.5 7B)
- Start chatting!
WP-Agentic-Admin uses Service Worker technology to keep the AI model loaded in memory, even when navigating between WordPress admin pages. This provides several benefits:
- Instant Access: Once loaded, the model is immediately available - no waiting
- Faster Navigation: Switch between admin pages without reloading the model
- Multi-Tab Support: Multiple browser tabs share the same model instance
- Better Performance: Reduced memory usage compared to loading separate instances
When you first load Agentic Admin in Chrome or Edge:
- The model downloads to browser cache (~1.2GB for Qwen 3 1.7B, ~4.5GB for Qwen 2.5 7B, one-time)
- A Service Worker registers and loads the model into GPU memory
- The model stays loaded as long as you have a Agentic Admin tab open
- Navigate away and back - the model is still there!
| Browser | Service Worker Mode | Fallback Mode |
|---|---|---|
| Chrome 113+ | ✅ Fully Supported | ✅ Available |
| Edge 113+ | ✅ Fully Supported | ✅ Available |
| Safari | ❌ Not Supported* | ✅ Auto-fallback |
| Firefox 120+ | ✅ Available |
*Safari's Service Workers cannot access WebGPU. The plugin automatically falls back to page mode, where the model loads in the page context.
When Service Worker mode is active, you'll see a "Persistent" badge in the model status area, indicating that your model will survive page navigations.
Current Status: 14 abilities, 4 workflows
| Ability | Description | Type |
|---|---|---|
cache-flush |
Flush WordPress object cache | Write |
db-optimize |
Optimize database tables | Write |
transient-flush |
Clear expired transients | Write |
revision-cleanup |
Remove old post revisions | Write (destructive) |
rewrite-flush |
Flush WordPress rewrite rules | Write |
rewrite-list |
List all registered rewrite rules | Read-only |
site-health |
Get comprehensive site health information | Read-only |
error-log-read |
Read recent entries from debug.log | Read-only |
cron-list |
List scheduled WordPress cron events | Read-only |
plugin-list |
List all installed plugins with status | Read-only |
plugin-activate |
Activate a specific plugin | Write |
plugin-deactivate |
Deactivate a specific plugin | Write (destructive) |
core/get-site-info |
Get WordPress site name, URL, version, and language | Read-only |
core/get-environment-info |
Get PHP version, database server, environment type | Read-only |
| Workflow | Steps | Purpose |
|---|---|---|
site-cleanup |
Cache flush → DB optimize (conditional) → Site health | Full maintenance routine |
performance-check |
Site health → Error log (conditional) | Quick diagnostic |
plugin-audit |
Plugin list → Site health | Review installed plugins |
database-maintenance |
DB optimize → Cache flush | Database optimization |
Diagnosing Errors:
You: My site is showing a white screen
AI: I'll check your error logs for recent issues.
[Executing: error-log-read]
I found the problem. There's a fatal error in the "broken-plugin" plugin:
"Fatal error: Call to undefined function xyz() in /plugins/broken-plugin/main.php"
Would you like me to deactivate this plugin?
Performance Issues:
You: The site feels slow lately
AI: Let me check a few things.
[Executing: site-health]
[Executing: db-optimize]
I found some issues:
- Your database has 15 tables that need optimization (now fixed)
- Object cache is not configured
- You have 3 inactive plugins that could be removed
Adaptive Tool Chains:
You: something is broken
AI: [Executing: site-health]
[Executing: error-log-read]
I found several PHP errors in your error log related to the "broken-plugin" plugin.
The most recent error is: "Fatal error: Call to undefined function..."
Would you like me to deactivate this plugin?
| Document | Description |
|---|---|
| AI Fundamentals | Understanding LLMs, WebGPU, quantization, and browser-based AI |
| ARCHITECTURE.md | Design principles and ReAct loop architecture |
| CLOUDFEST_HACKATHON.md | Feature roadmap and hackathon priorities |
| Abilities Guide | How to create new abilities |
| Workflows Guide | How to create multi-step workflows |
| Third-Party Integration | Extending with custom plugins |
| Testing Guide | Unit tests and E2E browser test suite |
Abilities - Atomic operations (list plugins, flush cache, etc.) Workflows - Pre-defined multi-step sequences triggered by keywords ReAct Loop - AI-driven tool selection based on observations and reasoning
wp-agentic-admin/
├── wp-agentic-admin.php # Main plugin file
├── uninstall.php # Clean uninstall (multisite-aware)
├── includes/
│ ├── functions-abilities.php # Public API: wp_agentic_admin_register_ability()
│ ├── class-abilities.php # Ability registration orchestrator
│ ├── class-admin-page.php # Admin page & assets
│ ├── class-settings.php # Plugin settings (model selection, etc.)
│ ├── class-utils.php # Utility helpers (cache invalidation, etc.)
│ └── abilities/ # Individual PHP ability files
│ └── shared/ # Shared helpers (plugin-helpers.php)
├── src/extensions/
│ ├── App.jsx # Main React app (Chat + Abilities tabs)
│ ├── sw.js # Service Worker for model persistence
│ ├── index.js # Entry point
│ ├── abilities/ # Individual JS ability files
│ ├── workflows/ # Workflow definitions (site-cleanup, etc.)
│ ├── components/ # React UI components
│ ├── services/ # Core services (ReAct agent, orchestrator, etc.)
│ │ └── __tests__/ # Automated tests
│ └── utils/ # Logging utilities
├── build-extensions/ # Compiled assets
└── docs/ # Documentation
cd wp-agentic-admin
npm install
npm run buildnpm test # Run unit tests (43 tests)
npm test -- react-agent # Run specific test file
npm run test:watch # Watch mode
# Ability tests — tool selection accuracy via local Ollama (auto-installs on macOS)
npm run test:abilities -- --file tests/abilities/core-abilities.test.jsAbility tests send prompts to a local Ollama instance running Qwen 3 1.7B and verify the model selects the correct tool. No browser or WebGPU needed — runs in ~20s.
The project also includes an E2E browser test suite that validates the full pipeline (user message → LLM reasoning → tool selection → tool execution → response) using Chrome DevTools MCP. See tests/TESTING.md for details.
This project includes a .claude/ folder with project rules, MCP servers, and pre-built skills for Claude Code contributors:
| Skill | Usage | Description |
|---|---|---|
/release |
/release 0.10.0 |
Version bump all files, commit, push, and create GitHub release |
/test |
/test |
Run lint, unit tests, and ability tests in sequence |
/new-ability |
/new-ability cache-warm |
Scaffold PHP + JS ability files with registration and test case |
/update-docs |
/update-docs |
Sync all documentation with current codebase state |
/pr |
/pr |
Create a PR with branch naming, pre-flight checks, and template |
/issue |
/issue add backup-check |
Create a GitHub issue via interview with codebase inspection |
The .mcp.json at the project root configures WordPress-specific MCP servers (wp-devdocs, wp-blockmarkup, wp-playground) that are automatically available to all contributors using Claude Code.
Client-Side:
- Runtime: WebAssembly & WebGPU
- AI: WebLLM with Qwen 3 1.7B (default) or Qwen 2.5 7B
- UI: React
- Chat: ReAct loop with adaptive tool selection
- Persistence: Service Worker mode keeps model loaded across navigation
Server-Side:
- Plugin: PHP 8.2+
- Protocol: WordPress Abilities API
- No Third-Party AI Services: The AI model runs entirely in your browser - no data is sent to external AI providers
- No Data Collection: Your site data stays between your browser and your WordPress server
- Permission-Based: All abilities respect WordPress capabilities
- Confirmation Required: Destructive actions require explicit approval
WP Agentic Admin aims to make WordPress site management accessible through natural language while maintaining privacy and security. We're building:
- Privacy-First AI - No external APIs, no data collection, fully local execution
- Extensible Architecture - Third-party plugins can add custom abilities via the WordPress Abilities API
- Smart Reasoning - ReAct loop to bridge natural language with technical operations
- Community-Driven - Open development with transparent roadmap and contributor recognition
See CLOUDFEST_HACKATHON.md for our feature roadmap and future plans.
- GitHub Discussions: Ask questions, share ideas, and connect with other users
- WordPress Slack: Join the
#agentic-adminchannel for real-time discussion - Issue Tracker: Report bugs or request features on GitHub Issues
We welcome contributions from developers, designers, testers, and documentation writers! Here's how to get involved:
- 🔧 Code: Implement new abilities, improve the ReAct agent, or fix bugs
- 📝 Documentation: Write guides, tutorials, or improve existing docs
- 🧪 Testing: Test with real WordPress sites and report issues
- 💡 Ideas: Propose new abilities or workflows in GitHub Discussions
- 🌍 Translation: Help translate the plugin into other languages
- Read our Contributing Guide for detailed guidelines
- Check our Code of Conduct
- Review the architecture documentation:
- ARCHITECTURE.md - ReAct loop design
- Abilities Guide - Implementing abilities
- Workflows Guide - Creating workflows
- Pick an issue labeled
good-first-issueor propose your own contribution
- Branch from
main- All PRs should target themainbranch - Follow WordPress Coding Standards - We use WordPress PHP and JavaScript standards
- Write tests - New abilities should include tests when applicable
- Document your changes - Update relevant documentation
- Disclose AI assistance - If you used AI tools, mention it in your PR description
- Keep PRs focused - One feature or fix per PR for easier review
For detailed guidelines, see CONTRIBUTING.md.
Thank you to all our contributors! See the GitHub Contributors page for the full list.
Want to see your name here? Check out our Contributing Guide!
GPL-2.0-or-later
- WordPress AI Team for the Abilities API
- WebLLM for browser-based LLM inference
- CloudFest Hackathon 2026 for the initial development sprint