Smart Local AI adds three AI features to WordPress without a cloud AI service: semantic related posts, AI alt text for images and personalized recommendations for every visitor. The machine learning runs on-device, in the browser, with Transformers.js and WebGPU or WebAssembly. Your posts and images are never uploaded to an AI service, there is no OpenAI or Google API key to buy, and there is no per-request cost.

Version 1.2 brings a new app-style settings screen: a menu on the left, short tabs instead of long pages, and a live preview on the right that shows the related-posts widget, the recommendations or an alt-text suggestion exactly as your changes will look, before you save.

At a glance

Three modules: RelevantFlow (semantic related posts), AltGenius (AI alt text generator) and PersonaFlow (personalized content and product recommendations). Turn each on or off on its own.
Where the AI runs: in the logged-in editor’s browser, never on your server and never in the cloud. Visitors only receive pre-computed results.
Any content type: posts, pages, WooCommerce products or any public custom post type. Every page recommends items of its own type, so product pages show products and blog posts show posts.
Cost: free. No API key, no account, no usage fees.
Privacy: results are saved in your own WordPress database. PersonaFlow follows your cookie banner through the WP Consent API.
Works with: block and classic themes, page caching plugins and CDNs, shared hosting, multisite and WordPress Playground.

Who it is for

Blogs and publishers who want related posts that understand meaning, not just tags
WooCommerce stores that want “you may also like” product recommendations without sending shopper data to a third party
Privacy-conscious sites (GDPR, public sector, health, education) that cannot send content to external AI APIs
Anyone improving image SEO and accessibility with descriptive alt text
Agencies that need one plugin that works the same on every client site, with no API keys to manage

Why site owners choose Smart Local AI

No API keys, no subscriptions – zero recurring AI costs, however many posts or images you have
Private by design – the AI reads your content in your own browser; only small results reach your database
Works on shared hosting – no Python, no GPU and no AI libraries on the server
Fast for visitors – lists are prepared in advance and loaded after the page, so they work with page caching
See it before you save – every display setting has a live preview, with a desktop and phone view
Guided setup – a setup wizard checks your browser, shows each model’s size before downloading, and lets you try every feature on your own content first

RelevantFlow: semantic related posts

Most related-posts plugins match tags, categories or keywords. RelevantFlow understands meaning. It turns each post into a sentence embedding with the all-MiniLM-L6-v2 model and recommends content that is genuinely similar, even when two posts share no tags at all.

Semantic similarity instead of keyword overlap: synonyms and related ideas match
Choose the content type to index: posts, pages, products or a custom post type
Each page shows items of its own type; switching type keeps the old index, so switching back is instant
Cards, list or grid layouts, with or without images and excerpts
Colour presets (follow theme, light, dark, minimal) or your own colours and corner radius
Placed under the content automatically, inside any element with a CSS selector, or anywhere with the [atlas_ai_relevantflow] shortcode
Leave out old posts, and fall back to shared categories and tags for posts that are not indexed yet
The index is built in your browser in batches; new and updated posts are flagged for re-indexing automatically

AltGenius: AI alt text generator

Missing alt text hurts image SEO and accessibility (WCAG). AltGenius writes descriptive alt text with an image-captioning model that runs in your browser, so your photos never leave your site.

Two models: ViT-GPT2 (fast, about 250 MB) or Florence-2 (more detailed, about 460 MB)
Suggests alt text automatically when you upload an image, or on request in the Media Library
Block editor panel with a Generate button for every image in a post
Minimum confidence: suggestions the model is unsure about are not shown
Optional text before and after each suggestion for consistent, brand-aware wording
You always review a suggestion before it is saved
Bulk alt text for the whole Media Library is part of Smart Local AI Pro

PersonaFlow: personalized recommendations

PersonaFlow recommends content to each visitor based on what they actually read, the way streaming services recommend shows. It is off until you turn it on.

30 behaviour signals, such as page views, reading time, scroll depth, read completion, site searches, likes, dislikes and bookmarks
A taste profile for each visitor, computed on your own server from RelevantFlow’s embeddings
New visitors see related items; personal picks appear after a few signals
Context-aware: product pages recommend products, articles recommend articles
WooCommerce signals: product views, add to cart, purchases, reviews and wishlists
Show recommendations instead of related posts, next to them, or with the [atlas_ai_personaflow] shortcode
Optional like, dislike and save buttons on every recommendation
Analytics for any date range: actions, visitors, signals and the most engaging content
Data is kept for 90 days by default (you choose) and never leaves your server

Privacy and GDPR

Smart Local AI shows no cookie banner of its own. It works with the one you already have:

WP Consent API support: with Complianz, CookieYes, Cookiebot, Real Cookie Banner, Cookie Notice or any other consent plugin that supports the free WP Consent API, PersonaFlow records only visitors who allow the category you choose (marketing by default). It starts the moment a visitor accepts and stops, deleting the visitor ID, when they withdraw consent.
Other consent tools can send one documented browser event, or answer a filter for page views counted by the server.
Everything is declared: a Consent and privacy screen lists every item PersonaFlow stores, where and for how long. The same list goes into your consent plugin’s cookie declaration and into Settings Privacy Policy guide.
Data requests: logged-in users’ PersonaFlow data is included in WordPress’s Export Personal Data and Erase Personal Data tools.
Post text and image pixels used by the AI never leave the editor’s browser. Usage data is sent to us only if an administrator opts in.

Every outside connection is listed under “External services”.

How the on-device AI works

The AI engine (ONNX Runtime for WebAssembly) ships inside the plugin. The open-source models are downloaded from the Hugging Face Hub only when you start the download, in the setup wizard or in a module’s settings:

RelevantFlow: Xenova/all-MiniLM-L6-v2, about 23 MB
AltGenius: Xenova/vit-gpt2-image-captioning, about 250 MB, or onnx-community/Florence-2-base-ft, about 460 MB

The browser keeps each model in its cache after the first download, so it loads instantly next time (this needs HTTPS). WebGPU is used when the browser supports it (Chrome and Edge 113+); otherwise the plugin uses WebAssembly. The work runs in Web Workers, so the editor stays responsive.

Built for AI assistants

Every REST endpoint and setting describes itself. With AtlasMCP, the MCP server for WordPress, assistants such as Claude, ChatGPT or Cursor can check the index, change settings, rebuild related posts or read PersonaFlow analytics, with the same permission checks as the dashboard.

Smart Local AI Pro

Smart Local AI Pro adds:

Several content types at once, for example blog posts and products, each with its own index, number of items, heading and visitor profile
Bulk alt text for the whole Media Library
77 extra PersonaFlow signals: reactions, star ratings, shares, session context and deeper WooCommerce events
Visitor controls: hide a post, mute or block an author, dismiss a category
Extended analytics: engagement funnel, WooCommerce funnel, devices and sessions

Pro plugs into the same settings screen and setup wizard. The free plugin keeps working fully without it. Questions? Contact us.

External Services

This plugin connects to the following external services. Your posts, images and visitor data are never sent to any of them.

Hugging Face Model Hub

AI models are downloaded from the Hugging Face Model Hub (https://huggingface.co/ and its file download servers) by your browser, only when you start a download (in the setup wizard, the module settings, or when you first use a module), and are then kept in the browser’s Cache Storage. Like any download, Hugging Face sees the downloading browser’s IP address. If you enter a Hugging Face access token for Florence-2, it is sent to Hugging Face with the model download. The ONNX Runtime WebAssembly engine that runs the models ships with the plugin and is not downloaded from a CDN. The following models are used:

RelevantFlow: Xenova/all-MiniLM-L6-v2 (~23 MB) – sentence transformer for content embeddings
AltGenius (ViT-GPT2): Xenova/vit-gpt2-image-captioning (~250 MB) – image captioning model
AltGenius (Florence-2): onnx-community/Florence-2-base-ft (~460 MB) – detailed image captioning model

These models are open source (Apache 2.0 or MIT). No user content, images or site data is uploaded to Hugging Face; the connection is download-only.

AtlasAiDev plugin list and WordPress.org plugin information

When an administrator opens Smart Local AI > Plugins, your server downloads the list of AtlasAiDev plugins from GitHub (https://raw.githubusercontent.com/atlasaidev/plugins/main/plugins.json) and their ratings and install counts from the WordPress.org Plugins API (https://api.wordpress.org/plugins/info/1.2/). Both are cached for 24 hours. No site data is sent.

AtlasAiDev promotions (opt-in only)

Only if an administrator has opted in to AtlasAiDev usage data (see below), your server checks a small JSON file on GitHub Gist (https://gist.githubusercontent.com/) for current AtlasAiDev offers, at most every 12 hours, and shows them as dismissible notices to administrators. No site data is sent. Without opt-in this file is never requested.

GitHub Terms of Service · GitHub Privacy Statement

AtlasAiDev usage data (opt-in only)

Only if an administrator opts in (on the last step of the setup wizard, or in the admin notice), your server sends non-sensitive usage data once a week to AtlasAiDev (https://track.atlasaidev.com/): WordPress, PHP and plugin versions, which modules are enabled, content counts and setup progress. To include the site’s public IP address in this report it calls https://icanhazip.com/. Nothing is sent without opt-in, and you can opt out at any time.

AtlasAiDev Privacy Policy

Smart Local AI – Related Posts, AI Alt Text & Personalized Recommendations

About this plugin