Agent Builder

⬇️ Install in WordPress (no coding needed)

1. Download agent-builder.zip. It is always the latest version.
2. In your WordPress admin, go to Plugins → Add Plugin → Upload Plugin, choose the file, and click Install Now.
3. Click Activate. Agent Builder opens a short setup wizard to connect your AI provider.

Updates then arrive in your normal WordPress update screen.

This download is the self-hosted build of Agent Builder: the same plugin as the code in this repository, packaged with its libraries and an updater that checks agentic-plugin.com. When Agent Builder is listed on WordPress.org you will also be able to install it from Plugins → Add Plugin by searching for "Agent Builder".

⚠️ Don't use GitHub's green Code → Download ZIP button. That is the source code, not an installable plugin: it is missing required libraries and unzips under the wrong folder name.

Version: 4.4.0

Free Agent Builder WordPress plugin.

• Plugin slug: agent-builder
• License: GPL-2.0-or-later
• Requires: WordPress 6.4+, PHP 8.1+
• Docs / product site: https://agentic-plugin.com/documentation/
• Community agents: https://agentic-plugin.com/community-agents/

Give AI agents real WordPress work. Hand a job to an agent, walk away, and get pinged only when it needs you. Low-risk steps run on their own; anything that needs your judgement waits for one-click approval.

What it does

• Tasks: assign a job in plain English. It runs in the background, pauses under Waiting on you when a step needs approval, then resumes.
• Site Brief: a read-only scan of your site turned into a ranked list of approve-to-act jobs.
• Routines: run an agent on a schedule or when something happens on the site, with test run, pause and a 20-run history.
• Agent profiles: a name, title, avatar and standing instructions for each agent. Pin, hide, duplicate, or export and import as a template.
• Skills: save a conversation as a skill, teach a task by demonstration, or invoke a skill with /skill-name.
• Approval rules: plain-English rules ("Ask me first before publishing anything"), plus allow once, for this task or always.
• Safety: risk levels on every tool, automatic backups, a hash-chained audit log and a kill switch.
• MCP and WebMCP: connect Grok Bot, Claude Desktop, Cursor or VS Code. Visitors' AI browser agents can use low-risk tools.
• Bring your own model: OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Kimi, Cohere or local Ollama.

12 agents included free

Agent	Role
Content Writer	Researches, writes, edits, and formats blog posts and pages
SEO Optimizer	Audits on-page content and proposes keyword and meta improvements
Site Health Sentinel	Continuously checks performance, database health, and security alerts
Support Triage	Summarizes customer comments, reviews form submissions, and drafts replies
WordPress Assistant	Onboards new users and helps troubleshoot core settings
Assistant Trainer	Builds new specialized AI agents from a plain-English job description
Agent Orchestrator	Deploys assistants as frontend chat widgets, admin launchers, or background jobs
Editorial Director	Plans editorial calendars and coordinates publishing workflows
User Assistant	Manages member outreach, onboarding, and role-based permissions
Skills Assistant	Discovers and imports community skills to teach agents new capabilities
Storefront Assistant	Helps visitors browse your WooCommerce catalog and build a cart, including via WebMCP
AI Radar	Checks how visible your site is to AI crawlers and assistants (robots.txt, llms.txt, schema.org) and fixes gaps with your approval

More at Community Agents.

Install from source (developers)

1. Clone into wp-content/plugins/agent-builder (or symlink).
2. Optional document tools: composer install --no-dev
3. Optional rebuild admin React: npm ci && npm run build Pre-built assets already ship in build/.
4. Activate Agent Builder in WordPress.

Project structure

Path	What's there
includes/	Core PHP: agent registry, tool executor, risk levels, audit log, REST API, admin menu
library/tools/	Every tool an agent can call — one self-contained class per directory (get_description(), get_parameters(), execute())
library/agents/	The 11 bundled agents — see library/README.md for the four-file format and how to build your own
admin/	Classic PHP admin screens (non-React)
src/	React admin surfaces (Dashboard, Safety Center, Settings, etc.), built via @wordpress/scripts into build/
templates/	Frontend chat widget and modal markup
tests/	PHPUnit suite (tests/unit/) — committed to git, excluded from the WordPress.org zip
library/prompt-tests/	The prompt-test catalog and its developer docs (most_popular_prompts.md, PROMPT-TESTING.md) — ships in the zip, so wp agent prompt-test works on every install

Adding a new tool means creating a directory under library/tools/, declaring its risk level (includes/class-risk-level.php documents the tiers and the reasoning behind each floor), and registering it in the relevant agent's abilities.json.

Development

• npm run start — watch mode for the React admin (src/)
• npm run build — production build into build/
• npm run lint:js / npm run format:js — @wordpress/scripts lint/format for src/

PHP tests

A real PHPUnit suite runs against a live WordPress core test install (WP_UnitTestCase), not a mocked sandbox:

composer install
bin/install-wp-tests.sh <db-name> <db-user> <db-pass> [db-host] [wp-version]
composer test

bin/install-wp-tests.sh needs a local MySQL to provision a test database against, and svn to pull the WordPress core test library — it's the standard WP-CLI plugin-scaffold installer, so any WordPress plugin dev environment already has both. It downloads WordPress core + the wordpress-develop test library into $WP_CORE_DIR/$WP_TESTS_DIR (default: /tmp/wordpress, /tmp/wordpress-tests-lib) and creates the test database — safe to re-run.

To run a single test file or method:

./vendor/bin/phpunit tests/unit/test-risk-level.php
./vendor/bin/phpunit --filter test_dangerous_tools_are_never_silently_allowed

Coverage priorities live under tests/unit/: Risk_Level (enforcement matrix + BASELINE_RISKS floor), Audit_Log_Integrity (hash-chain tamper detection), Approval_Queue (create/approve/reject/expire), Tool_Executor (the full risk-gate flow — allow/confirm/queue/block, plus that a non-readonly tool triggers a table backup before it runs), and Abilities_Manifest (effective-risk resolution). tests/unit/test-tool-risk-floor-coverage.php is a standing regression test: it fails the moment a new non-readonly tool ships under library/tools/ without either a get_risk_level() override or a Risk_Level::BASELINE_RISKS entry.

No test makes a real outbound HTTP call — tests/helpers/MockWPFunctions.php intercepts wp_remote_*() via the pre_http_request filter for any test that needs one.

Prompt tests (real LLM calls)

Alongside the hermetic PHPUnit suite there's a prompt-testing harness — PHPUnit for the agents rather than for the code. It replays a ranked catalog of 54 real-world prompts (library/prompt-tests/most_popular_prompts.md) against the bundled agents on a live install and writes a markdown report.

wp agent prompt-test --dry-run                             # resolve and validate, no LLM calls
wp agent prompt-test --user=admin --agent=seo-optimizer
wp agent prompt-test --user=admin --rank=1-10 --max-cost=0.50

Unlike composer test this makes real LLM calls against your configured provider and costs real money, and prompts that write really write — so it's deliberately not part of composer test or CI, and a bare invocation runs nothing until you select a subset. The catalog and the parser are covered by the PHPUnit suite, so a bad tool name or agent slug fails for free in milliseconds rather than partway through a paid run.

The catalog ships inside the plugin and is treated as read-only; the first edit copies it to wp-content/agentic-knowledge/prompt-tests/, so your additions survive plugin updates and nothing writes inside the plugin directory.

Assistant Trainer can also drive this from chat — running prompts, then proposing the tools or skills an agent turns out to be missing. It proposes; you approve. Full instructions, every flag, and how to add prompts: library/prompt-tests/PROMPT-TESTING.md.

Releases

GitHub tags match plugin versions (v3.3.0, etc.). Downloadable ZIPs for production installs are published from the product site and WordPress.org once listed.

Support

• WordPress.org support forum
• https://agentic-plugin.com/support/