Every feature, and all of it free

A four-step wizard turns one API token into a working chatbot.

Activation opens a setup wizard that runs four steps against the Cloudflare API: verify the token and resolve your account, deploy the Worker and create the vector index, downgrade to a minimal operational token, and index your content. If your token covers several Cloudflare accounts the wizard asks which one to use. Every step is idempotent, so re-running setup repairs a half-finished install rather than duplicating resources.

The powerful setup token is never stored, and is deleted when setup finishes.

The setup token exists only in the browser and in the request bodies of the setup calls, and is never written to the database. Once provisioning succeeds the plugin mints a replacement token scoped to just Workers Scripts Edit and Vectorize Edit on the one account, stores that encrypted, and then asks the setup token to delete itself from your Cloudflare account. What is left on your site is the least privilege needed to redeploy the Worker and rebuild the index.

Chat, ingest, delete and health endpoints, deployed to your own workers.dev subdomain.

The plugin uploads a bundled Worker script named after a hash of your site URL, so one Cloudflare account can host chatbots for many sites without collisions. It binds Workers AI, your Vectorize index, a plain-text config blob and a Durable Object rate limiter, and serves from your workers.dev subdomain. Settings changes redeploy the Worker rather than reaching for KV, so there is no extra moving part to go wrong.

Serve the chatbot from chat.yoursite.com instead of workers.dev.

If the site’s zone is on the same Cloudflare account you can attach a custom hostname from the settings screen, and the plugin creates the Worker domain for you. It also tries to add a zone rate-limiting rule as a second layer of abuse protection. Both are optional: the workers.dev URL works on its own, and sites whose DNS is not on Cloudflare are fully supported.

Answers are drawn from your own content, with citations, and it declines when it does not know.

Each question is embedded and matched against your indexed content, keeping the three closest chunks above a confidence threshold. Every chunk carries the URL of the page it came from, so the bot can link the visitor to the source. When retrieval turns up nothing relevant the bot says so and points at your human support link instead of inventing an answer, and if retrieval fails outright it degrades to a polite generic reply rather than an error.

Posts, pages, any public post type, plus chatbot-only knowledge entries.

Pick the post types the bot should read: posts and pages by default. Content is stripped, split into roughly 1,500-character chunks and embedded. A per-post "exclude from the chatbot" checkbox keeps individual items out. Chunk identifiers are derived from the post, so when a post gets shorter or is deleted the stale vectors are removed instead of lingering and being cited months later.

Answers the bot should know that you do not want published on the site.

A dedicated Chat Knowledge post type holds material that feeds the index but never renders publicly: answers to questions nobody wrote a page about, internal policy wording, the phrasing you want used for a delicate topic. Alongside it, a Site Facts box injects always-true statements straight into the system prompt, which is far more reliable than hoping retrieval surfaces them.

Edits re-index shortly after saving; a daily sweep catches anything missed.

Saving a post queues a re-index moments later. Unpublishing or deleting removes its vectors straight away. A daily sweep re-indexes anything modified since the last run, and the initial bulk index runs as a batched queue (a handful of posts per tick) so large sites do not trip PHP timeouts. "Sync content now" forces the whole thing from the settings screen.

Zero third-party JavaScript, streamed replies, and no framework to load.

The front-end widget is hand-written and dependency-free: no chat library, no framework, nothing pulled from a CDN. It streams the reply token by token as the model produces it, renders a deliberately narrow subset of Markdown (bold, inline code, fenced code, safe links) with everything escaped first, and opens from a floating button with an optional greeting bubble and up to six suggested questions.

Pick a colour; contrast is calculated for you and previewed live.

Choose the widget colour from your theme’s own palette in one click, or set your own. Rather than trusting that choice, the plugin computes the relative luminance of the colour and derives text, bubble and accent colours that clear WCAG AA contrast against it. The settings screen previews both the closed button and the open panel using the real widget markup, so what you see is what visitors get.

Origin allowlist, per-IP rate limiting, and a prompt built to resist injection.

The chat endpoint is public, and it spends your Workers AI allocation, so it is locked down: browser requests are only accepted from your own domain, each IP is limited to 20 messages a minute by a Durable Object counter, and conversation history is capped in both message count and message length. The system prompt is hardened against the usual injection attempts (role spoofing, "ignore your instructions", requests to reveal the prompt) and treats retrieved content as untrusted data rather than instructions.

Daily probes, an admin notice when something breaks, and a second model to fall back on.

Workers AI models do get retired, and a chatbot that quietly stopped answering is worse than one that never launched. A daily health check exercises the Worker and its bindings and raises a WordPress admin notice the moment something fails. The Worker itself tries a configured fallback model before giving up, and both the chat and fallback model identifiers are editable, so you can move off a deprecating model without waiting for a plugin release.

Diagnostics, forced re-sync, index rebuild and a local reset.

The settings screen carries the tools you need when something is off: run diagnostics against the live Worker, force a full content re-sync, rebuild the vector index from scratch (required if you change the embedding model), or reset the plugin’s local state and start over. Uninstalling removes the plugin’s own data and deliberately leaves your Cloudflare resources alone, so nothing you paid for or built on is deleted behind your back.