Approval rules that understand plain English
- Approval rules now read your rule: before an agent acts, each rule is checked against that specific action. “Never delete users” no longer pauses a read-only summary, “Allow automatically: creating draft posts” lets drafts through without asking, and a “Never” rule blocks the action and tells you which rule stopped it. High-risk actions still always need your OK. Each check sends a short, redacted description of the action to your configured AI provider.
- DeepSeek and other “thinking” models: tasks and routines no longer fail on models that reject a forced tool choice.
- Privacy: full AI request bodies are no longer written to your server’s error log when a provider call fails.
- Accurate error reports: Site Brief and Site Health Sentinel count real PHP errors only, not log lines that merely mention the word “fatal”.
- Cleaner deactivate and uninstall, a tool cache that refreshes on every update, and a tidier step list in Tasks.
Connectors and team runs in Pro, a security update & Gemini fixes
- Pro 4.2.0 — Team runs: give a task to a Lead agent with “Run as team” and 2–6 specialists; the Lead plans, the specialists work in parallel, and you get one combined answer in Tasks. One new Team screen holds Team runs, Automations and Traces.
- Pro 4.3.0 — Connectors: add Stripe, PayPal, HubSpot, Notion, Canva or Zapier from a curated catalogue and sign in with OAuth. Tools from these services start as read-only and need your approval, and everything they return is treated as untrusted outside data.
- Pro 4.2.1 — Security update: connection secrets now use authenticated encryption and upgrade automatically; outbound requests to MCP servers are pinned to verified addresses; connector tools check the acting user’s permissions in scheduled runs too. Recommended for every Pro site.
- Gemini fixes (4.3.2–4.3.3): agents on Google Gemini or the hosted provider no longer fail when a tool has no parameters or returns a list, and the hosted provider now shows your remaining credits.
- Bundled agents refreshed (4.3.1): versions bumped where behaviour changed, and every description now matches what the agent actually does — see the marketplace.
Skills you can teach, result cards & live activity
- Save as skill: turn any chat conversation into a reusable skill from the message menu — the draft opens for your review before it is ever used.
- Teach a task: start a recording, show an agent the steps, then stop; Agent Builder drafts a skill from what it watched.
- Run with
/: type/and a skill's name in chat to run it; new skills are shared with all your agents by default (changeable in Skills). - Result cards: posts, files and lists an agent creates come back as cards with Open and Edit links instead of plain text.
- Live activity: a Working pane shows each step as it happens — how long it took and whether it needed approval — and background tasks show the same steps in Tasks.
- Hardened chat: one shared markdown renderer for every chat surface, and voice dictation now works on any HTTPS site without enabling audio replies.
Approval rules & Routines
- Approval rules: write plain-English rules such as "Ask me first before publishing anything" or "Never delete users", scoped to one agent or all. Rules only ever tighten approvals, and "ask first" always wins over any saved permission.
- Allow for this task: a new option on approval cards sits alongside Allow once, Allow this session and Always allow; the Approvals → Grants list shows your saved permissions so you can revoke them.
- Routines: Agent Builder → Routines lists every scheduled and event-triggered routine with its next run, the last 20 runs, Test run, Pause and Resume — plus a banner when WP-Cron looks stalled.
- Safer approvals: approved actions re-check arguments and risk and take a backup before they run, and proposals are stored durably instead of expiring.
- Claude 5.x: the Claude 5 family (Sonnet 5.5, Opus 5.5, Fable 5.1 and more) joins the model picker.
- Pro 4.1.1 / 4.1.2: AI Radar listener throttled;
wp agent infocompatible with 4.2.
Tasks, notifications & agent profiles
- Tasks: assign work to any agent from Agent Builder → Tasks; it runs in the background and only comes back when it needs an approval or has a result. Active / Waiting on you / Done tabs, cancel and retry, and an inbox count in the admin bar.
- Notifications & daily digest: administrators get a daily activity email (switch to instant, or turn it off, under Settings → Security); each user can opt out on their profile.
- Agent profiles: give each agent a display name, title, avatar and standing description — pin, hide and reorder agents, duplicate one, or export it as a template to import on another site.
- Background run health: a Site Health check watches WP-Cron health, stuck runs and approvals waiting over a day; runs survive PHP time limits and resume after you approve.
- Full run history: every run is recorded with its owner, steps, tokens and result, and linked from the audit log.
- MCP clients: works with Grok Bot, Claude Desktop, Cursor and VS Code over MCP.
- Pro 4.1.1 / 4.1.2: AI Radar listener throttled;
wp agent infocompatible with 4.2.
Security & dashboard fixes
- Link hardening: chat replies can no longer render
javascript:,data:or other unsafe link schemes, and link URLs are attribute-escaped across every chat surface. - Dashboard fix: the Agent Builder dashboard no longer renders blank, and the site-wide Basic/Advanced switch is back at the top right of every screen.
- Clearer states: before a provider is connected, links explain "connect a provider first" instead of a false "Administrator access required" screen.
- Approval cards: cards now show the post title, status and type you are approving, scroll into view, and no longer display escaped characters.
- Emergency Stop visibility: the dashboard shows when Emergency Stop has halted all agents, and stale "Agentic → Deployments" references now read "Agent Builder → Publish".
Site Brief, Prompt Tests & a hardened, safety-first 4.0
- Site Brief: a read-only scan that turns your site into a ranked list of approve-to-act jobs, each attributed to the agent that raised it — nothing changes until you approve.
- Prompt Tests: replay a catalog of real owner requests against your agents; see what passed, what stopped for approval and what it cost, and get proposed improvements for your approval.
- Safety Center: one screen for risk inventory, approval gates, per-agent tool scopes, tamper-detection status, and an emergency stop.
- Agent-Ready Score & WebMCP Bridge: check and improve how discoverable your site is to AI agents (WebMCP is off by default).
- Hardened throughout: hash-chained, tamper-evident Activity log; a stronger approval write-path; and a consistent
agent_builder_prefix across constants, options, tables, hooks and capabilities. - All three editions in sync: WordPress.org, self-hosted and Pro all ship at 4.0.0.
Model capabilities, Knowledge Wiki & admin polish
- Model capabilities layer: provider-aware feature detection for safer tool/model pairing.
- Knowledge Wiki (OKF): free local markdown knowledge for agents; Pro Vector Store remains for large corpora.
- Admin UX: React dashboard/settings, Deployment tabs, scheduled tasks, and documentation updates.
- Also archived: 3.2.3 3.2.2 3.2.1 3.2.0
The Agent Library release
- Contextual AI launchers: one-click prompts now appear across wp-admin — Plugins, Media Library, Users, Comments, and the Dashboard — opening chat pre-seeded with a task-appropriate prompt. Editable before sending, never auto-sent, dismissible per user.
- DB-backed Agent Library: agents now install from a signed library — marketplace purchases are verified with Ed25519 publisher signatures before they can run.
- Declarative manifest agents: build agents with
agent.json+abilities.json— no PHP required, and community agents run sandboxed. - Autonomous work: scheduled tasks and event listeners let agents act on cron schedules or WordPress hooks, hands-free.
- Agents can fetch a human: the escalation tool ships in free; Pro fans requests out to Slack and email.
- WordPress Abilities, inbound & outbound: the WP Abilities tab splits into Inbound (what your agents expose to WordPress core AI) and Outbound (native hooks your agents can invoke).
- Per-agent MCP endpoints + cache tools: every agent gets its own MCP endpoint, plus a
manage_cachetool that detects and purges popular page-cache plugins. - Bigger library: 261 built-in tools, 26 skill packs, and 8 bundled agents — content-writer, seo-optimizer, site-health-sentinel, support-triage and more.
- Pro 3.0.0 alongside: channels (WhatsApp, Slack, Telegram, SMS, email and more), durable workflows, connectors, budget caps and the tamper-evident audit ledger.
User Assistant (Pro) & the connector wave
- User Assistant (Pro): a security-aware agent for managing the people on your site — reviews registrations, privileged and inactive accounts, failed-login signals, and contributor authorship. Account-security actions (lock account, force password reset) run only after your confirmation.
- Google Workspace connector: agents can list and create Calendar events, search and read Drive files, and read and append Docs — with credentials encrypted at rest.
- CRM connector: HubSpot and Salesforce behind one interface — agents can create, update and find contacts, open deals, and log notes.
- Signed outbound webhooks: POST signed JSON to Zapier, Make, n8n or any endpoint — every delivery carries an HMAC signature.
- Facebook Messenger & Instagram DM channels: talk to visitors on Meta's platforms through the same agent runtime.
Usage dashboard rebuilt
- Live cost trends: the Usage page is now a live React dashboard — summary cards, per-agent cost trend chart, and a breakdown table with a 7/30/60/90-day period selector.
- Also in 2.18: Pro now ships in 11 languages, and a dedicated Settings → License tab manages your key, activation status, and renewal in one place.
Observability & evaluation
- End-to-end run traces: every agent run is captured step-by-step — prompt, tool call, result, tokens, cost, and latency — so you can see exactly what happened and why.
- Eval harness: define test prompts and expected behaviours per agent, run them on demand, and see a pass/fail regression diff before rolling out a prompt or model change.
Governance & cost control
- Budget hard-caps: per site, agent, or user — spend is throttled, then stopped, so a runaway agent can never surprise you.
- Egress allowlists + DLP: restrict which domains agent tools may call, and redact PII from outbound request bodies automatically.
- Tamper-evident audit ledger: hash-chained, append-only logging with CSV, JSON, and SIEM-webhook export.
- Role-scoped permissions: map WordPress roles to the agents and scopes each role may use.
Channels, voice & human handoff
- New channels: Discord, Telegram, Microsoft Teams, SMS (Twilio), and a generic authenticated inbound webhook — all routed into the same agent runtime.
- Voice telephony via Twilio: multilingual speech, barge-in, and voicemail-to-agent — one agent across web, WhatsApp, Slack, and phone.
- Human handoff: pause an agent, take the visitor to a live operator in the handoff console, then resume.
Marketplace, Durable Workflows & MCP
- Marketplace: install signature-verified, sandboxed agents in two clicks — with ratings, reviews, semantic search, and version pinning. Publish your own trained agents for sale.
- Durable Workflows: build multi-step agent workflows in a visual node-graph builder — triggered by webhooks, schedules, WordPress hooks, or channel messages, with conditional branching, parallel steps, and human approval pauses.
- Outbound MCP client: agents can consume tools from external MCP servers, with per-tool risk gating through the same approval path as built-in tools.
- Slash Commands moved to Pro: the full command registry, per-agent toggles, and admin UI now live in the Pro plugin.
Consistent vertical navigation everywhere
- Unified admin layout: Activity, Deployment, Knowledge, Approvals, and Tools all now share the same left vertical nav introduced in Settings — easier to scan, easier to navigate, consistent everywhere.
Account tab
- At-a-glance account status: A new Account tab in Settings shows your current plan and which AI providers have an API key configured — displayed as a masked connection status, never shown in full. Free users see a direct link to explore Pro.
Settings redesign & global agent instructions
- Vertical navigation + search: Settings now uses a left-side nav with an instant search box — find any option without scrolling through tabs. Mirrors the layout of familiar desktop AI apps.
- General tab: Set site-wide context that applies to every agent — how agents address admins and visitors, shared background knowledge ("Instructions for Agents"), and a global chat font and accent colour.
Agent Creation Wizard
- "Train an Agent" wizard: A guided 5-step React flow — name, persona, provider/model, autonomy level, optional tools — lets you create and activate a new AI agent in minutes. Accessible from the Dashboard with a single click. No config files, no code.
Settings grouped into clusters
- Basic / Security & Limits / Advanced: Settings tabs are now organised into labelled groups so common options are front and centre and developer-only tabs (APIs, Endpoints) stay tucked away in Advanced mode.
Live Activity dashboard card
- Auto-refreshing React card: The Dashboard activity card is now live — it refreshes automatically and adds a Week / Month period toggle, replacing the static weekly-stats snapshot.
Simplified admin navigation & redesigned Dashboard
- Plain-English menu: "Deploy" renamed to "Publish", "Logs" renamed to "Activity". A new Basic / Advanced interface toggle hides developer tools (Skills, APIs, Endpoints) so the menu stays approachable for non-technical users.
- New Dashboard: Getting Started checklist, estimated usage stats, and quick actions trimmed to the most common tasks.
Multi-agent orchestration, local memory & quick-start templates
- Multi-agent orchestration: Team-lead agents can delegate subtasks to other active agents sequentially, with configurable depth, fan-out, and cost/token budgets. An Editorial Director template that orchestrates your specialist agents ships in the box.
- Local conversational memory: Agents recall relevant context from a user's past conversations — stored entirely on your server, no external cloud service (opt-in: Settings → Security → Local Memory). Respects chat retention windows and GDPR erasure requests.
- Quick-start templates: Four one-click agent templates: Support Triage, SEO Optimizer, Site Health Sentinel, and the Editorial Director team lead.
- WordPress Abilities readiness: A banner and outbound abilities list under Tools → WP Abilities shows your site's readiness for the WordPress native AI runtime (WP 6.9 / 7.0) with actionable guidance.
- Accessible UI: All blocking alert()/confirm() dialogs replaced with accessible toast notifications and keyboard-navigable modals (Esc to cancel, focus trap, screen-reader roles).
Cleaner Quick Start & smarter first-run experience
- Quick Start page: Clearer registration copy, corrected consent wording, and a skip link that lists all supported AI providers on a single line — less friction, more clarity.
- Focused first-run menu: When no LLM provider is configured, the admin menu shows only Quick Start and redirects all other Agent Builder pages there — no dead ends on a fresh install.
- AI Radar 1.2.0: Improved weekly scan prompt, smarter event listeners for robots.txt changes, and updated bot definitions.
Exception handling & security logging across all critical paths
- Hardened runtime: Every critical execution path — tool calls, chat streaming, approval queue, job processor — is now wrapped in try/catch with structured exception logging. Unexpected errors are caught and recorded rather than silently failing or producing a blank screen.
- Granular Security Log: All agent operations now emit structured Security Log entries including start, success, and failure states — easier to diagnose issues without turning on WP_DEBUG.
- Agent update fix: The one-click agent update now correctly handles zip archives that wrap files in a subdirectory (the most common GitHub release format).
Privacy improvements & cleaner activation
- No silent phone-home: The plugin no longer contacts agentic-plugin.com when it is activated — zero external requests fire without explicit administrator consent.
- Non-intrusive setup banner: Activation now shows a dismissable admin notice instead of forcing a redirect to the sign-up page, so WP-CLI and headless installs work correctly.
- Single-request sign-up: The sign-up flow now returns the license key alongside the API key in one request, removing a background HTTP call that happened silently after registration.
- Direct to chat: After completing setup, you land straight in the WordPress AI Assistant — no extra clicks.
New security tools, Forms Builder improvements & freemium UX
- 4 new Security Assistant tools: Publicly accessible files audit, file permissions checker, PHP compatibility scanner, and hosting plugin conflict detector — added after a real-world security incident highlighted these gaps.
- WordPress AI Assistant: Now gives actionable answers including code snippets, settings paths, and step-by-step instructions — the previous read-only restriction was over-broad and has been corrected.
- Forms Builder: Native form styling, new form list tool, and supervised mode fixes. Supports Contact Form 7, WPForms, Gravity Forms, and Fluent Forms.
- WP-CLI: New
wp agent update <slug>command to update agents from the terminal. - Freemium UX: Clearer quota error messages with direct links to buy credits. Dashboard upgrade nudge. 7-day conversation history available on the free tier.
Pro terminology & privacy update
- Privacy: Agent update checks are now strictly opt-in. The plugin will never contact agentic-plugin.com without explicit administrator consent — you'll see a clear consent screen on your first visit to the Agents page.
- Terminology: The plugin consistently uses "Pro" throughout the admin interface, matching the branding on the website and marketplace.
New AI models, Video Assistant & Logs redesign
- New models: Claude Sonnet/Opus/Haiku 4.x, Gemini 2.5 Pro/Flash/Flash-Lite, Grok 3 and Grok 3-mini added to the model picker.
- Video Assistant: New premium agent that generates video clips from text prompts with full post-production — trim, stitch, add background music, and generate captions. Powered by Google Veo via the Agentic service.
- Unified Logs: Audit Log, Conversations, and Security Log redesigned as a single tabbed interface with period filters (Today / Last 7 Days / Last 30 Days).
- Per-role usage limits: Set independent daily AI quotas for anonymous visitors, subscribers, contributors, authors, and editors.
Free Agentic AI included & privacy improvements
- Free AI out of the box: 50 AI chat queries per day at no cost, no API key required — powered by the Agentic AI service. Set up in seconds from the welcome screen.
- Privacy: IP addresses anonymised by default. GDPR data exporters and erasers expanded to cover all plugin data.
- Explicit consent: A clear sign-up screen explains exactly what data is shared before you connect to the Agentic AI service.
3 core free agents + safe Trainer (unlimited custom), daily quotas & Conversations
- Post-gap baseline (see strategy): exactly 3 core free agents (WordPress Assistant, Content Writer, Assistant Trainer) + safe Assistant Trainer for unlimited custom + 5 templates. (Earlier entries listed 10 including Pro/marketplace specialists; free is now precisely 3 + Trainer per F0–F4.)
- Daily usage quotas: Free tier enforces per-user daily limits on chat queries, image generation, and RAG lookups — with clear user-facing messages and a link to buy credits.
- Admin notifications: Email alert when users hit their daily quota.
- Conversations page: Browse and review all user chat sessions from the admin panel.
Redesigned admin menu & agent configuration
- New admin menu: Chat, Agents, Deploy, Tools, Skills, and Integrations — reorganised for clarity.
- Agent memory: Agents can retain information between sessions. Configure retention TTL per agent.
- Response style presets: Choose Balanced, Concise, Detailed, Technical, or Friendly for each agent.
- Suggested prompts: Customise the prompt quick-picks shown to users in each agent's chat.
Context-aware chat, Google Sheets & file uploads
- Context-aware chat: Chat widgets automatically inject the current page title, URL, and referrer into the conversation — agents know where the user is without any extra config.
- Google Sheets channel (Pro): Connect agents to read from and write to Google Sheets spreadsheets using a service account.
- GDPR consent gate: Optionally require visitor consent before the public chat widget activates.
- Image & file uploads: Users can attach images and files directly in the chat interface.
- Multimodal (vision): Agents can analyse image attachments using any vision-capable AI model.
GDPR consent gate & Costs tab
- GDPR consent (REST): The REST API is gated — AI requests cannot be sent until the visitor accepts the consent banner.
- Costs tab: View current AI model pricing across all providers directly inside your WordPress admin — no more tab-switching.
- Branded email alerts: Security events (failed logins, file changes) now trigger styled admin emails.
Knowledge Files & RAG via WP-CLI
- Knowledge Files: Agents declare knowledge files in their configuration and receive the content in their system prompt automatically — give agents reference documents, style guides, or FAQs they can always draw on.
- WP-CLI RAG commands: Manage your knowledge base from the terminal —
wp agent rag status|train|upload|delete|search. - Document formats: PDF, plain text, and Markdown files all supported.
Plugin update checker, deployment tabs & modal widget
- Auto-update checker: Get Agent Builder updates directly in your WordPress Plugins screen — no WP.org directory needed (Pro).
- Deployment tabs: Configure where each agent is deployed — admin chat, public shortcode, Gutenberg block, admin sidebar, or floating modal.
- Modal widget: A floating "chat" button that opens a full agent chat panel on any page — zero-code public deployment.
Plugin Assistant with Gutenberg blocks & MU plugin support
- Plugin Assistant improvements: Generates complete WordPress plugins from a plain-English description including React Gutenberg blocks, WP-CLI commands, REST API endpoints, and MU plugin variants.
- Multi-turn chat: Agents maintain context reliably across long conversations — better follow-up and iteration.
- Assistant Trainer hardening: More accurate agent generation with improved system prompts.
Vision model per provider & Ollama DB storage
- Vision model: Configure a dedicated vision-capable model per AI provider — e.g. use GPT-4o for vision and GPT-4o-mini for text to keep costs down.
- Ollama endpoint: Ollama server URL stored per-provider in the database — no more editing config files.
Anthropic (Claude) integration & write operations ledger
- Anthropic provider: Full Claude integration — tool calling, streaming, response normalisation. Default model set to Claude Sonnet.
- Write operations ledger: Every non-read-only tool call is logged with timestamp, mode (supervised/autonomous), and invocation context (chat/schedule/hook/CLI). 30-day retention.
- AI Radar: Preview robots.txt and llms.txt changes before writing — read-only preview tools bypass the approval queue.
Post scheduling, password reset tool & Gutenberg sidebar
- Post scheduling: Agents can now schedule posts for future publication — not just immediate publish.
- Force password reset: Security agents can flag a WordPress user account for a forced password change on next login.
- Gutenberg sidebar: Provider configuration warning now shown in the sidebar before you type anything — clearer feedback when the provider isn't set up.
- Per-shortcode styling: Each
[agentic_chat]shortcode can set its own theme, colours, and height independently.
Health tab redesign & per-service credit tracking
- Health tab: Completely redesigned — live service status indicators for every connected provider.
- Per-service credits: View your remaining AI credit balance broken down by service (chat, RAG, images, TTS).
- TTS toggle: Enable or disable text-to-speech output globally from the admin settings.
- Tab reordering: Drag settings tabs into any order and reset to defaults with one click.
MCP integration — connect Claude Desktop, VS Code & Cursor
- MCP server: Agent Builder exposes your agents as an MCP server. Connect Claude Desktop, VS Code Copilot, Cursor, or any MCP-compatible client and use your site's agents as tools directly from your editor or AI assistant.
- Per-agent MCP servers: Enable MCP on a per-agent basis with configurable tool scopes.
- WP-CLI MCP command:
wp agent mcp-connectto test the connection from the terminal.
Gutenberg sidebar, event triggers & email templates
- Gutenberg editor sidebar: Chat with any agent without leaving the editor — get writing help, SEO suggestions, and content ideas while you're writing.
- Event trigger management: Browse and manage all active WordPress hook listeners from a dedicated UI.
- Email templates: Customise the HTML email templates sent by agent-triggered notifications.
WP-CLI agent execution
- WP-CLI tool: Run agent tasks directly from the command line — ideal for server-side automation, cron jobs, and CI/CD pipelines.
Safety & trust layer — risk levels and approval workflows
- 5-tier risk system: Every agent tool is classified as None / Low / Medium / High / Extreme risk. You decide which risk levels require manual approval before execution.
- Approval queue: High-risk actions pause and queue for administrator review — agents cannot execute destructive operations without human sign-off.
- Tamper protection: Agent configurations are integrity-checked on load to detect unauthorised modifications.
WhatsApp channel integration
- WhatsApp (Pro): Connect your agents to WhatsApp via the Meta Cloud API. Visitors and customers can chat with your AI agents directly through WhatsApp — no app to install on your server.
White-label branding
- White-label: Hide or replace "Agent Builder" branding in the chat interface — ship AI assistants under your own agency or product brand.
Granular permissions & proposal/review system
- Granular permissions: Control exactly what each agent can read and write — content, files, plugins, settings, emails, and more — independently per agent.
- Proposal/review system: Agents in supervised mode describe what they're about to do and wait for your approval before making any changes.
- Automatic backups: Files and database tables are backed up before any agent write operation.
- Write permissions off by default: New agents start in read-only mode for safety.
Real-time event listeners
- Event listeners: Agents react to WordPress hooks in real time — post published, comment added, user registered, plugin activated, and dozens more. No polling required.
- Async task handling: Agent responses in autonomous mode are processed asynchronously so they never block page loads.
Scheduled tasks & Security Monitor
- Scheduled tasks: Agents can run recurring jobs on any cron schedule — daily SEO audits, weekly security scans, monthly reports.
- Security Monitor: Daily autonomous scan for failed logins, outdated plugins, modified files, and suspicious admin accounts.
Agent Builder — initial stable release
- Core AI agent runtime — create, configure, and run AI agents inside WordPress.
- Multi-provider support: OpenAI, Google Gemini, xAI, Mistral, Meta Llama, and Ollama.
- Tool system: content creation, SEO analysis, security scanning, file management, and more — extensible by developers.
- Full audit log of every agent action.
- WP-CLI, WP REST API, and Gutenberg block integrations.
Ready to try it?
Download the free Agent Builder plugin — no license required to get started. Upgrade later if you need Pro features.
Download free Agent Builder