Description

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.

Features

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: AI inference runs in your browser – your prompts and content are never sent to an AI service. A few abilities look up public data (CVE databases, WordPress.org checksums, web search); all are listed under External services
Zero Server Costs: No GPU infrastructure needed – computation happens on the client
WordPress Abilities API: Natively integrates with WordPress’s official Abilities API
Natural Language Interface: Describe problems in plain English, get intelligent solutions

Requirements

WordPress 6.9+ (includes the Abilities API)
PHP 8.2+
Modern browser with WebGPU support (Chrome 113+, Edge 113+)

External services

This plugin runs AI locally in your browser by default, and your prompts and chat content are never sent to an AI service unless you explicitly enable the external LLM provider below.

Separately from AI inference, some abilities query public data sources to do their job: a security scan checks your plugin versions against CVE databases, a checksum verification compares your files against WordPress.org, and a web search sends your query to a search engine. Every external request the plugin makes is listed here:

Model source for the local engine (site-owner configured) — Only when an administrator sets a model source and loads a model.
The local engine runs a language model in the administrator’s browser. The model is not part of the plugin: model files are over 1 GB. The plugin contains no model download address. An administrator enters where models are downloaded from under Settings Model source, and until then the local engine is off. Nothing is downloaded until an administrator then selects a model and clicks Load Model. The browser fetches the files directly from the configured source (the WordPress server makes no requests for them), and no prompts, admin data, or telemetry are sent. Files are cached in the browser after the first download.

To use the models published by the MLC-AI project, enter these addresses. No account or API key is needed:

Model weights URL: https://huggingface.co/mlc-ai/
Model library URL: https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/

You can also host the same files yourself and enter your own https addresses. Use the same layout: each model’s weights in <weights URL>/<model ID>/resolve/main/, and the compiled model libraries in <library URL>/<WebLLM version>/ (the plugin adds the version it needs, currently v0_2_80).
Hugging Face terms: https://huggingface.co/terms-of-service — Privacy: https://huggingface.co/privacy
GitHub terms: https://docs.github.com/en/site-policy/github-terms/github-terms-of-service — Privacy: https://docs.github.com/en/site-policy/privacy-policies/github-general-privacy-statement

External LLM provider (user-configured) — Only when you switch the engine from “Local” to “Remote” in settings.
When enabled, chat messages, tool descriptions, and tool results are sent through this plugin’s REST proxy (/wp-json/agentic-admin/v1/llm-proxy/) to the OpenAI-compatible endpoint URL you configure (e.g. Ollama, LM Studio, vLLM, OpenAI, Groq, Together). You choose the endpoint; the plugin does not preselect or default to any third-party provider. No data is sent until you save an endpoint and start a chat in Remote mode.

DuckDuckGo HTML search — Only when the optional web-search ability is invoked by the assistant.
The user’s search query is sent to https://html.duckduckgo.com/html/ over GET. No WordPress user data is sent.
DuckDuckGo terms: https://duckduckgo.com/terms — Privacy: https://duckduckgo.com/privacy

NVD CVE database (NIST) — Whenever the security-scan ability is invoked.
security-scan includes a plugin vulnerability check, so the names and versions of your active plugins are sent to https://services.nvd.nist.gov/rest/json/cves/2.0 to look up known CVEs. No user data, post content, or credentials are sent.
NVD terms: https://nvd.nist.gov/developers/terms-of-use — Privacy: https://www.nist.gov/privacy-policy

MITRE CVE API — Only when the security-scan vulnerability check finds a CVE identifier and follows up on it.
The CVE ID (e.g. CVE-2024-12345) is sent to https://cveawg.mitre.org/api/cve/ for details. No WordPress user data is sent.
MITRE terms: https://www.cve.org/Legal/TermsOfUse — Privacy: https://www.cve.org/Legal/PrivacyPolicy

WordPress.org plugin checksums — Only when the optional verify-plugin-checksums ability is invoked.
Installed plugin slugs and versions are sent to https://downloads.wordpress.org/plugin-checksums/ and https://plugins.svn.wordpress.org/ to verify file integrity against the official WordPress.org distribution.
WordPress.org policies: https://wordpress.org/about/privacy/

WordPress.org core source (SVN) — Only when the optional verify-core-checksums ability finds a modified core file and diffs are requested.
The original copy of that one file is fetched from https://core.svn.wordpress.org/tags/{version}/ so the plugin can show you a diff against your local version. Only the WordPress version number and the core file path are sent, both of which are public information; no site data is transmitted. The checksum list itself comes from WordPress core’s own get_core_checksums() function.
WordPress.org policies: https://wordpress.org/about/privacy/

Source code and build process

Agentic Admin is fully open source under GPL-2.0-or-later. The complete, human-readable source — including the JavaScript/React sources behind the compiled assets in build-extensions/ — is maintained in a public repository:

https://github.com/pluginslab/wp-agentic-admin

The files under build-extensions/ are generated from the sources in src/ with @wordpress/scripts (webpack). They contain only JavaScript and CSS: the bundles (index.js, sw.js, and code-split chunks) and the stylesheets. No WebAssembly, binaries, or other compiled artifacts are distributed with the plugin. To regenerate them from a checkout:

npm install
npm run build

There is no build step for the PHP. The WebLLM engine is bundled into the plugin from its npm package. The AI model weights and their compiled model libraries are loaded at runtime from the provider documented under External services above; these are large provider-hosted model files (over 1 GB), not part of the plugin code.

Screenshots

The Agentic Admin chat tab in wp-admin, mid-conversation. The model has just answered a question about installed plugins by calling the plugin-list tool locally — full ReAct trace (user question, thought process, tool call, answer) visible. — The Agentic Admin chat tab in wp-admin, mid-conversation. The model has just answered a question about installed plugins by calling the `plugin-list` tool locally — full ReAct trace (user question, thought process, tool call, answer) visible.

First-run model download in progress. The Qwen 3 1.7B weights (~1.2 GB) are fetched from the configured model source — once per browser, cancellable, cached for subsequent sessions. — The Agentic Admin chat tab in wp-admin, mid-conversation. The model has just answered a question about installed plugins by calling the `plugin-list` tool locally — full ReAct trace (user question, thought process, tool call, answer) visible.

Installation

Upload the agentic-admin folder to the /wp-content/plugins/ directory
Activate the plugin through the ‘Plugins’ menu in WordPress
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!

Reviews

There are no reviews for this plugin.

Contributors & Developers

“Agentic Admin” is open source software. The following people have contributed to this plugin.

Marcel Schmitz

Translate “Agentic Admin” into your language.

Interested in development?

Browse the code, check out the SVN repository, or subscribe to the development log by RSS.

Changelog

0.11.0

Renamed: Plugin is now “Agentic Admin”. Text domain “agentic-admin”, function prefix agentic_admin_*. WordPress.org submission-ready.
Removed: feedback system, WebMCP bridge, voice input, file editing (write-file), and AI content generation (content-generate), plus three low-value abilities (backup-check, opcode-cache-status, disk-usage). Agentic Admin is a site health and maintenance assistant: it does not edit files or generate content. Removed code remains available in the project’s git history.
Security: blocked sensitive-column reads (user_email, user_pass) in query-database to prevent reconnaissance attacks (#166). Hardened query length cap and read-only verb gate.
Security: escaped output in functions-abilities.php (#121). Documented direct-DB-call rationale in db-optimize and database-check.
Docs: corrected the External services section. The CVE lookups were attributed to a plugin-vulnerability-scan ability that does not exist; they are made by security-scan. Added the previously undisclosed core.svn.wordpress.org fetch used by verify-core-checksums to build diffs. Replaced two absolute privacy claims that the section’s own list contradicted.
Compliance: removed deprecated load_plugin_textdomain() call, prefixed uninstall.php globals, fixed nonexistent Domain Path header, excluded CLOUDFEST_HACKATHON.md from distribution build.
New: ability manifest as single source of truth for which abilities register, with a agentic_admin_enabled_abilities filter for selective override. PHP authoritative for PHP-backed abilities; JS adds back the JS-only ones.
Fixed: thinking-disable setting now actually suppresses the streaming UI even when Qwen ignores /nothink (#181).
Fixed: Settings panel is now a real tab (not a stateful cog toggle). No more lingering Chat-tab underline or stale-state regressions (#197).
Fixed: Activate/Deactivate buttons on plugin-list rows now flip to the inverse action after a successful click (#179).
Fixed: AI model no longer preloads on every wp-admin page, deferred until the user opens the sidebar for the first time (#116).
Improved: post-tool summarization is brief (no re-listing items the user already sees in the tool result UI).
Improved: ChatInput keyboard handling simplified (Space inserts a space, no push-to-talk hijacking).
Removed: the local knowledge base (in-browser embeddings, vector store, and the codebase-index, code-search, schema-extract, wp-api-extract, docs-extract, and codebase-extract abilities). The plugin no longer loads Transformers.js from a CDN or ships any WebAssembly module.
Security: read-file now refuses files that can hold credentials, keys, or salts (wp-config.php, .env, private keys, database dumps, and any file defining the auth keys, salts, or DB password) instead of relying on redaction.
Security: wp-config-list now has its own PHP backend that reads constant names with the PHP tokenizer and never reads the values of credentials, keys, or salts.
Security: removed the security-scan salts check, which read the auth salt and could never fail.
Changed: AI Connectors are detected through the AI Client registry, as on the core Connectors screen, instead of reading API-key options.
Fixed: plugin, must-use plugin, and content paths are derived from the plugin’s own location and WordPress functions instead of WP_PLUGIN_DIR, WPMU_PLUGIN_DIR, and WP_CONTENT_DIR. Core checksum verification no longer miscounts files when the content directory is moved. The debug log path follows WordPress’s own error_log setting.
Fixed: read-file and uploads-scan resolve the site root with get_home_path(), so subdirectory installs work.
Security: the external LLM proxy no longer relays the provider’s response verbatim. Each streamed event is decoded and re-encoded as escaped JSON; anything else is dropped.
Changed: “Tested up to” is declared only in readme.txt.
Fixed: the model notice names where the model is downloaded from and its real size, instead of saying no data is sent to external servers.
Changed: the plugin no longer contains any model download address. The site owner sets the model source under Settings Model source (the MLC-AI addresses are listed under External services), and the local engine stays off until it is set. WebLLM’s built-in model list is removed at build time.
Fixed: .well-known scanning resolves via get_home_path() instead of ABSPATH, so subdirectory installs scan the real site root.
Removed: 7 stale tab references and 6+ stale docs files (FEEDBACK-DEV.md).
Tests: 96 unit tests passing, plus the new manifest test suite (7 cases), index test suite (6 cases), and react-agent regression tests (3 cases for the per-call state cleanup fix).

0.10.0

CloudFest Hackathon 2026 release: 78 PRs merged, 42+ abilities shipped by 11 contributors
New: Plugin Abilities Platform, third-party plugins can auto-register abilities with the AI assistant
New: Voice input via local Whisper model, all audio stays on-device
New: In-browser RAG knowledge base with vector embeddings computed client-side
New: Settings UI with GPU detection, context window tuning, and knowledge base toggle
New: External AI provider support (Ollama, LM Studio, OpenAI, any compatible endpoint) via PHP REST proxy
New: WebMCP bridge for external AI agents via navigator.modelContext
New: “Check if Hacked” multi-step security workflow
New: Gutenberg editor sidebar and admin bar chat integrations
New: 28+ new abilities including web-search, security-scan, plugin vulnerability scanning, read-file, write-file, query-database, content-generate, and more
New: E2E test runner mirroring browser ReAct loop
New: CI/CD via GitHub Actions (PHP lint, JS lint, unit tests, build check)
New: Feedback system with thumbs up/down stored locally
New: Distribution zip for WordPress Playground compatibility
Improved: WCAG 2.2 AA accessibility compliance
Improved: GPU capability detection with f32 precision fallback
Improved: Fuzzy plugin name matching for activate/deactivate actions

0.9.5

New: Split Gutenberg Integration into three separate goals — Block Editor Chat Window (Goal 7), Update Post Content Ability (Goal 8), WASM Dev Docs Abilities (Goal 9)
New: Goal 9 aligned with existing feature/wasm-sqlite-devdocs branch work (4 abilities, 11 tests)
New: GitHub issues created for Goals 6-9 (#32-#35)
New: Voice + Context Recording separated as Goal 1.5 (stretch)
New: WebMCP Integration added as Goal 6 (stretch)
Improved: CLOUDFEST_HACKATHON.md — all goals now follow consistent format, code examples stripped from stretch goals

0.9.4

Updated: WebLLM upgraded from 0.2.80 to 0.2.82
New: /playground Claude Code skill for spinning up WordPress Playground with the plugin mounted and activated

0.9.3

New: ONBOARDING.md — step-by-step hackathon guide with project tour and role-specific instructions
New: /issue Claude Code skill for creating GitHub issues via interview with codebase inspection
Improved: CLOUDFEST_HACKATHON.md simplified — removed duplicated role descriptions and project summary, now links to ONBOARDING.md
Improved: README.md updated with /issue skill in Claude Code integration table

0.9.2

New: AI Fundamentals glossary with ~60 terms linked back to source chapters
New: Chapter 13 — “Designing Abilities” guide for Writers and UX contributors
New: End-to-end walkthrough in Chapter 8 tracing a message through the full system
New: Testing guidance section in Chapter 11 for LLM Testers writing tool selection tests
Improved: Trimmed overly theoretical content from Chapters 1, 2, and 5
Improved: Deduplicated hardware requirements — Chapter 3 is now single source of truth
Improved: INDEX.md updated with new “For Writers & Designers” section

0.9.1

Cleaned up: Removed stale version annotations (v1.0–v2.0) from comments across 11 source files
Fixed: model-loader.js comment incorrectly referenced Qwen2.5-7B instead of actual default Qwen3-1.7B
Fixed: ARCHITECTURE.md listed v0.7.0 as current version, updated to v0.9.0
Fixed: THIRD-PARTY-INTEGRATION.md referenced non-existent v1.4.1 for includeIf feature
Improved: README.md release example updated to current versioning

0.9.0

New: CloudFest Hackathon 2026 preparation with 28 GitHub issues across 7 contributor roles
New: CLOUDFEST_HACKATHON.md with sharpened goals, team role definitions, and demo flow
New: LICENSE file (GPL-2.0-or-later)
New: SECURITY.md with vulnerability reporting policy
New: GitHub labels for hackathon workflow (php, react-js, ai-ml, ux, docs, devops, testing, writers, ability, workflow, hackathon)
New: README.md header image from CloudFest pitch deck
Improved: CONTRIBUTING.md version requirement corrected to WordPress 6.9+
Fixed: readme.txt changelog version numbering (internal 1.x versions renumbered to public 0.x scheme)

0.8.0

New: .claude/ project config with CLAUDE.md rules, settings, and 5 contributor skills
New: /release, /test, /new-ability, /update-docs, /pr slash commands for Claude Code
New: .mcp.json with wp-devdocs, wp-blockmarkup, and wp-playground MCP servers
New: Chrome DevTools MCP plugin auto-enabled for E2E testing
New: Pre-configured permissions (safe commands allowed, destructive commands denied)

0.7.1

New: Ability metadata auto-loaded from JS source files — no static manifest to maintain
New: Test suite expanded to 18 tests covering all 14 abilities (was 8 tests covering 5)
Improved: JS lint fixes for test runner and helpers

0.7.0

Changed: Ability test runner now uses Ollama (local LLM server) instead of Puppeteer + WebGPU browser
Changed: Tests run in ~20s vs 5+ minutes — no browser, no WebGPU, no model download required
Removed: Puppeteer dependency and browser-based test harness
Removed: build:test-harness build step — npm run test:abilities runs directly via Node.js
Improved: Test runner auto-installs Ollama via Homebrew and pulls models on first run

0.6.0

New: 4 WP-CLI-inspired abilities for common maintenance tasks:
- transient-flush – Delete expired or all transients (like wp transient delete)
- cron-list – List scheduled cron events with overdue detection (like wp cron event list)
- rewrite-flush – Flush permalink rewrite rules (like wp rewrite flush)
- revision-cleanup – Delete old post revisions with preview mode (like wp post delete for revisions)
New: WordPress 6.9 core abilities support – chat wrappers for core/get-site-info and core/get-environment-info
New: Abilities browser now displays both Agentic Admin and WordPress core abilities
New: Smart confirmation for revision cleanup (preview/dry-run skips confirmation)
Improved: Better keyword separation to avoid triggering multiple abilities
Improved: Chat session persistence across page refreshes
Improved: Total of 14 abilities now available (12 custom + 2 core wrappers)
Fixed: HTTP method for destructive abilities (now uses POST as required by Abilities API)

0.5.0

New: Multi-step workflow engine for chaining abilities together
New: Built-in workflows: Site Cleanup, Performance Check, Plugin Audit, Database Maintenance
New: wp.agenticAdmin.registerWorkflow() API for third-party workflow registration
New: Workflow progress indicator in chat UI
New: Ad-hoc workflow creation from multi-intent messages
Changed: Default model upgraded to Qwen3-1.7B for better reasoning with native function calling (~1.2GB)
Improved: Confirmation dialog shows all workflow steps before execution
Improved: Automatic rollback of completed steps if later step fails

0.1.0

Initial release