Rapls AI Chatbot – Self-Hosted RAG & MCP Server

By rapls

★ 4.9/5 (10 ratings)70+ active installationsVersion 1.25.2

Answers from your own posts and pages, says so when it doesn't know, and stops at the spending limit you set. Works with OpenAI, Claude or Gemini.

Description

Add an AI chatbot that answers visitors from your own content — and admits when it doesn’t know. Rapls AI Chatbot searches your posts, pages, and knowledge base first, then replies in natural language, in the visitor’s own language. With Grounded Answers Only mode, if your content has no answer, the bot says so instead of making something up.

You can be live in about five minutes, for free, with no credit card. On first setup a guided “Start for free” panel hands you a free key from OpenRouter or the Google Gemini free tier, tests it, auto-selects a working model, and switches the chatbot on — no AI or API experience needed.

Watch the setup, start to finish:

No monthly SaaS fee — and no runaway bill. It runs on your own API key with no markup, and built-in Usage Control caps daily usage per visitor, so you can open the bot to the public without worrying about your API spend. A Monthly Cost Guard lets you set a hard budget, and Model Fallback keeps the bot answering on a lightweight model when your main model hits its quota. Conversations and keys stay on your own server.

Built for site owners, agencies, and developers who want control over the model, the data, and the cost. Connect OpenAI, Anthropic Claude, Google Gemini, or OpenRouter, or any OpenAI-compatible endpoint such as Alibaba Tongyi Qwen (DashScope) — useful where OpenAI and Gemini are hard to reach — and switch anytime.

Why site owners pick Rapls

  • Free to try, fast to launch. Guided onboarding gets you from install to a working chatbot in minutes, with a no-credit-card OpenRouter or Gemini key.
  • Grounded answers, or an honest “I don’t know.” RAG hybrid search grounds replies in your actual posts, pages, and knowledge base — and Grounded Answers Only mode keeps the bot from inventing answers your site doesn’t contain.
  • Your bill can’t run away. Self-hosted and BYOK with no markup, plus per-visitor daily caps (Usage Control) so public traffic can’t drain your API budget.
  • Replies that don’t sound like a robot. An optional AI-smell score reviews the bot’s Japanese replies for machine-sounding wording, so you can tune the tone. (Detection only — replies are never altered.)
  • No lock-in. Switch between OpenAI, Claude, Gemini, and OpenRouter whenever you want.
  • Speaks your visitors’ language. Automatic multi-language replies, so one bot serves an international audience.

What it does

  • Bring your own key: OpenAI, Anthropic Claude, Google Gemini, OpenRouter, or any OpenAI-compatible endpoint (Alibaba Qwen/DashScope, DeepSeek, Zhipu), switchable per site.
  • Site learning: indexes your posts and pages so the bot answers from your content.
  • Knowledge base: add Q&A or upload TXT/MD/PDF/DOCX files (large documents are split into parts automatically so they embed reliably), with priority over general answers.
  • RAG hybrid search: combines keyword and semantic retrieval for grounded replies.
  • Grounded Answers Only: an optional mode that makes the bot reply “I couldn’t find that” instead of inventing an answer your content doesn’t support.
  • Web search: lets the bot pull current information when configured.
  • MCP tools: exposes 5 Model Context Protocol tools, so agents such as Claude or ChatGPT can read and act on your site through conversation.
  • Usage dashboard: tracks conversations, messages, and API cost.
  • Usage Control: optional per-visitor daily caps so public traffic can’t run up your API bill.
  • Unanswered Questions: records what the bot could not answer, with one-click “Add to knowledge” and an AI-drafted FAQ page.
  • AI-smell score: an optional read-only score that flags machine-sounding wording in Japanese replies (detection only — replies are never altered).
  • Industry starter templates: one click applies a ready-made system prompt and preset questions for hotels, shops, clinics, professional services, or company sites.
  • Gutenberg block: drop the chatbot into any page or post.
  • Deep-link auto-send: a ?raplsaich_q= URL opens the chatbot and sends the question automatically, so end-of-article “suggested questions” become one-click conversations.

Turns visitor questions into content

The bot doesn’t just answer — it tells you what your site is missing. Questions it could not answer from your content are recorded on the dashboard, one click turns them into a knowledge entry, and another click drafts an FAQ page from the last 30 days of real visitor questions. Answer gaps become content, content improves the bot, and the same pages work for SEO and AI search (LLMO).

Self-hosted and private

Conversations and keys stay on your own WordPress install. You are billed by your AI provider directly, so cost is transparent and there is no markup. Visitors can delete their own chat history from the widget (optional), and retention is configurable — GDPR-friendly by design.

Free and Pro

Free — not a trial. The free version is the full product: chatbot, site learning, knowledge base, RAG hybrid search, Grounded Answers Only, MCP server (5 tools), a usage dashboard (conversations, messages, and cost), and per-visitor daily caps. Run it forever at no cost on a free-tier key.

Pro — for sites that run the bot as a business tool:

  • Lead capture automation — collect contacts from conversations, with Google Sheets and Slack sync.
  • Analytics — conversation, cost, and quality metrics at a glance.
  • Usage Control suite — role-based limits, per-user monthly credits with auto-reset, and a usage-control dashboard to grant and track credits, so members and guests each get a fair share of your API budget.
  • Extra MCP tools — product-search and analytics tools for AI agents, on top of the five in Free.
  • Context & embedding controls — reprocess embeddings and shape the context retrieved for each reply.
  • LINE integration (add-on) — connect the chatbot to LINE messaging, the dominant messenger in Japan.

Up and running in about 5 minutes (free, no credit card)

You do not need an API key or any AI experience to start. On first setup the plugin shows a “Start for free” panel with two no-credit-card paths: an OpenRouter free key or a Google Gemini free-tier key.

  1. Pick OpenRouter or Gemini.
  2. Click through to get a free key (about a minute) and paste it in.
  3. Press Test Connection — the key is validated and saved, a working free model is auto-selected, and the chatbot is switched on.

That’s it. Each option states its data-handling trade-off up front (the Gemini free tier may use submitted content to improve Google’s models), so you choose with eyes open. You can switch to your own OpenAI, Claude, or Gemini key at any time.

Learn more: Plugin details | Source code (GitHub)

See it in action

Watch the full setup — from install to first reply — in about 5 minutes, no credit card required:

See it answer from your own content, and how that differs from plain ChatGPT:

External Services

This plugin connects to the following external third-party services. No data is sent to any service until you configure an API key and enable the feature in the plugin settings. Each service requires the site administrator to create an account and obtain API credentials. By using these services, you agree to their respective terms and privacy policies listed below.

1. OpenAI (GPT models): AI Provider

Used when you select OpenAI as your AI provider. User messages and optionally site content are sent to generate AI responses.

2. Anthropic (Claude models): AI Provider

Used when you select Anthropic Claude as your AI provider. User messages and optionally site content are sent to generate AI responses.

3. Google (Gemini models): AI Provider

Used when you select Google Gemini as your AI provider. User messages and optionally site content are sent to generate AI responses.

4. OpenRouter: AI Provider

Used when you select OpenRouter as your AI provider. OpenRouter is a unified API gateway that routes requests to various AI models.

5. OpenAI-compatible endpoints (e.g. Alibaba Qwen/DashScope, DeepSeek, Zhipu): AI Provider

Used only if you select the “OpenAI-compatible” provider and enter a base URL. User messages, and optionally your site content (including indexed content for RAG embeddings), are sent to the endpoint you configure to generate responses or embeddings. Because the endpoint is chosen by you, please review the terms and privacy policy of the specific provider you connect to. For example, Alibaba Cloud Model Studio (DashScope), which powers Tongyi Qwen:

6. Google reCAPTCHA v3 (Optional)

Used only if you enable reCAPTCHA in the plugin settings for spam protection. The visitor’s IP address and interaction data are sent to Google for verification. By default Google’s script is loaded on every page that shows the chat; with “Load reCAPTCHA: Only when the chat is opened”, it is loaded only when a visitor opens the chat or clicks into its input.

7. LINE Messaging API (Pro Add-on, Optional)

Used only if you enable the LINE integration via the Pro add-on. Connects to the LINE Messaging API for chatbot-to-LINE messaging.

Cross-Site Embed

The plugin includes an optional embed loader script (embed-loader.js) for embedding the chatbot on external websites via an iframe. This script does not load any external CDN resources or third-party scripts. It creates an iframe pointing back to your own WordPress site, and all data processing occurs on your server.

Data Transmitted to External Services

  • User messages: Chat messages entered by visitors (sent to the configured AI provider only)
  • Site content (if Site Learning is enabled): Excerpts from your published posts/pages (sent to the configured AI provider)
  • Knowledge base (if configured): Custom Q&A entries you create (sent to the configured AI provider)
  • IP address (reCAPTCHA only): Sent to Google for spam verification

Data Storage

  • Conversation history: Stored locally in your WordPress database (can be disabled)
  • Visitor IP: Stored as SHA-256 hash (not plain text) for rate limiting
  • Retention: Configurable auto-deletion period (default 90 days)

User Controls

You can disable these features in the plugin settings:
* Conversation history saving
* Site content crawling/learning
* Google reCAPTCHA verification
* Web search

Installation

  1. Upload rapls-ai-chatbot folder to /wp-content/plugins/
  2. Activate via Plugins menu
  3. Go to AI Chatbot > Settings
  4. Follow the onboarding panel to start free with an OpenRouter or Google Gemini key, or paste your own OpenAI / Claude / Gemini key.
  5. Enable site learning or create knowledge base entries
  6. Insert Gutenberg block or enable sitewide display

Getting Started

  1. API Key: get a free OpenRouter or Google Gemini key from the onboarding panel, or your own key from console.anthropic.com, OpenAI, or Google AI Studio.
  2. Enable RAG: add site learning (auto-crawl) or create knowledge base entries.
  3. Customize: set bot name, avatar, welcome message, system prompt.
  4. Deploy: insert the Gutenberg block, paste the shortcode, or enable sitewide display.

👉 Plugin details: Rapls AI Chatbot

FAQ

My API key stops working every few days and I have to re-enter it

The API key (and the Pro license) is encrypted with a key derived from your WordPress security salts (AUTH_KEY / AUTH_SALT). If something changes those salts, the stored key can no longer be read. WordPress uses salt values kept in the database whenever the wp-config.php constants are missing, still set to “put your unique phrase here”, or reused across constants — and those database values can be regenerated without wp-config.php changing. Security plugins that rotate salts, a failing object cache, and restored database backups are the usual causes; you will normally be logged out of WordPress at the same time.

To make the stored keys independent of the salts, add a long random string to wp-config.php, above the “That’s all, stop editing!” line:

define('RAPLSAICH_ENCRYPTION_KEY', 'a-long-random-string-of-at-least-32-characters');

Existing keys are re-encrypted with it automatically the next time you open wp-admin, and they then survive any salt change. Keep the value backed up: losing it means re-entering your API key and license once.

Can I use it for free?

Yes — the plugin is free, and the onboarding panel connects a no-credit-card OpenRouter or Google Gemini free-tier key in about a minute. Free tiers have rate limits, so for production traffic you may want your own provider key.

Do I need an API key?

Yes. Rapls runs on your own API key so your data and cost stay under your control. If you do not have one, the plugin guides you to a free OpenRouter key or a free Google Gemini key on first setup — both need no credit card.

Will the bot make things up?

Not if you don’t want it to. Turn on Grounded Answers Only and the bot replies “I couldn’t find that on this site” whenever your content has no relevant answer, instead of answering from the model’s general knowledge.

Can visitors run up my API bill?

No. Usage Control caps daily usage per guest and per logged-in user on the server side, separate from rate limiting. Pro adds role-based limits and per-user credits.

Does it work with AI agents (MCP)?

Yes. The free version ships an MCP (Model Context Protocol) server with 5 tools, so AI agents such as Claude or ChatGPT can search your posts, read knowledge entries, and act on your site through conversation. Pro adds product-search and analytics tools on top.

Is my data private on the free options?

It depends on the provider you choose, and the plugin tells you before you pick. OpenRouter free models are served by various upstream providers, each with its own data-handling policy. The Google Gemini free tier may use your submitted content to improve Google’s models — if you do not want that, use a paid Gemini tier or another provider. Either way, your conversations and keys are stored on your own WordPress install, not on Rapls servers.

My Gemini key starts with “AQ.” instead of “AIza” — is that OK?

Yes. Google is moving Gemini API keys from the legacy “AIza” standard format to the new “AQ.” auth format, and Google AI Studio now issues “AQ.” keys by default. Rapls works with both. Google is retiring the standard format (unrestricted “AIza” keys are rejected from 2026-06-19, and all “AIza” keys from 2026-09), so if you still have an “AIza” key the plugin shows a notice with how to migrate. New keys created today are already “AQ.” keys and need no action.

Can I use multiple AI providers?

Yes. Configure multiple providers in Settings and switch between them. WordPress 7.0 Connectors API also supports unified key management.

Does it crawl my entire site automatically?

Yes. Site Learning is on by default: it learns your published posts, pages and other public post types once a day, and learns a post again when you update it. Turn it on or off, or run it right away, in Settings > AI Settings > Site Learning. Choosing which content types to learn and how often, excluding pages, managing the learned pages and learning WooCommerce product data are part of Pro.

Can I embed on external sites?

Yes. Configure cross-site embed in Display Settings. Use iframe or script loader.

How do I set up WordPress 7.0 Connectors API?

In Settings > AI Settings, Connectors UI appears if WP 7.0 is active. Register your Claude API key once; all Connectors-compatible plugins access it.

Is there conversation history?

Yes. Conversations are saved in your WordPress database, and the Data Management tab sets how many days they are kept (0 keeps them indefinitely) or turns saving off. The usage dashboard shows conversation and message counts. The screen for browsing individual conversations is part of Pro.

Troubleshooting: chat not appearing?

  • Verify API key is valid (test in Settings)
  • Check Gutenberg block or Display Settings (enable sitewide)
  • Review Security Diagnostics for rate limit, IP detection, or consent issues

More questions?

See Plugin details or WordPress.org Support

Screenshots

Changelog

1.25.2

  • Fixed: when only some page types were selected under Settings > Display Settings > Page Type Display, the chat still appeared on pages of other kinds, such as WooCommerce products, attachments and the 404 page. A new “Other pages” option covers them. It is on by default, so nothing changes until you turn it off. Thanks to @awmaeder for the report.

1.25.1

  • Fixed: since 1.25.0, some answers had page titles stuck to the end of sentences (“…to cut costs Heating.”). More learned pages now reach the AI with each question, and some models copied the title and source link of each page into the answer. The AI is now told that these only show where a passage comes from, and that it may link a page on a word inside the sentence but not after it. Thanks to @awmaeder for the report.
  • Added: “Remove page titles added to the end of sentences” under Settings > AI Settings > Site Learning > Links in Answers, off by default. If titles still appear, it removes a link to a source page at the end of a sentence, a bracketed source such as “(Source: URL)”, and a page title in 【】. Links inside a sentence and after a colon are kept.
  • Fixed: the Black Friday notice on the settings screen now ends on December 1, 2026, 20:00 JST (11:00 UTC), when the discount codes expire.
  • Changed: Rapls AI Chatbot now requires PHP 8.2 or later. PHP 7.4, 8.0 and 8.1 no longer receive security fixes. Sites on an older PHP version keep the version they have; WordPress does not offer this update until PHP is upgraded.

1.25.0

  • Added: the chat window’s own texts (buttons, notices and error messages) are shown in the visitor’s language. WordPress translates them only into the site language, and only when a language pack exists, so a German site without a German pack showed them in English, and with Response Language set to “Auto-detect” a visitor reading another language saw them in the site language. Translations for the 13 response languages are now built in: with “Auto-detect” they follow the visitor’s browser, otherwise the language you set. Texts you wrote yourself, such as the welcome message, are not changed. The translations are loaded only when they can be needed. Thanks to @awmaeder for the suggestion.
  • Added: “Reference Count” under Settings > AI Settings > Site Learning, so sites without Pro can choose how many sources (learned pages and knowledge base entries) each answer uses. A higher number helps when the same topic is spread over many pages.
  • Fixed: long questions gave worse answers than short ones. A word that appears on almost every page, such as the site or organisation name, counted as much as the word the question was about, so pages that only mentioned the name pushed out the pages that answered the question; on some sites the search never got past that word at all. Each word now counts by how rare it is on the site, words found on most pages are left out of the search, and the best matches are fetched first. Thanks to @awmaeder for the report.
  • Fixed: a high-priority knowledge base entry that shared a single word with the question could take the place of the page that answered it. Matches are now ordered by how much of the question they contain.
  • Fixed: each answer uses at least five sources, but only three learned pages were looked up, so the other places could only be filled with knowledge base entries.
  • Developer: the filters raplsaich_frontend_i18n_scripts and raplsaich_strings_language let add-ons add their own translations of chat window texts.

1.24.2

  • Fixed: on sites with five or more active knowledge base entries, answers could ignore the pages that matched the question. Every answer is built from up to five sources, and the knowledge base entries, which are added to every question whatever was asked, scored far higher than site pages and filled all five, so a page containing the exact name or term a visitor asked about never reached the AI. Adding a few large entries could make information that was found the day before disappear. Pages and entries that match the question now come first. Thanks to @awmaeder for the report and testing.
  • Added: Settings > AI Settings > Site Learning > “Knowledge Used on Every Question” chooses how the knowledge base shares those five places. “Use what matches the question first” (the default) keeps at least three for matches and fills the rest with knowledge. “Add it to what matches the question” uses all five for matches and adds the knowledge on top, which sends more text to the AI and costs more. “Use the knowledge first” keeps the previous behavior.
  • Changed: keyword and vector results are now combined by rank (Reciprocal Rank Fusion) instead of a weighted sum of their scores, so a page that matched the exact words of the question is no longer pushed out by pages that are only loosely similar in meaning.

1.24.1

  • Fixed: the “Links in Answers” option from 1.24.0 was only on the Site Learning screen, which appears only with Pro, so sites without Pro could not turn it on. It is now also under Settings > AI Settings > Site Learning. Both places change the same setting.

1.24.0

  • Added: an option to remove links to pages on your own site that do not exist. AI models sometimes make up URLs, especially for posts. When it is on, every link to your site in an answer is checked against the learned pages and WordPress (posts, pages, media, category and tag archives) before the answer is saved and shown; links to pages that are not found are removed and only their text is kept. Links to other sites are not changed. Turn it on under Site Learning > Learning Settings > Links in Answers; it is off by default. Thanks to @awmaeder for the idea.

1.23.1

  • Fixed: on themes that give every text box a minimum height (such as Cocoon on phones), the chat’s message box was several lines tall. It is one line again.
  • Fixed: with the iOS Safari keyboard fix on, the page showed through between the chat window and the keyboard on iPhone Safari. That space now uses the chat’s background color.
  • Fixed: “Bottom, clear of the keyboard buttons” did not add its space in Firefox on iPhone.

1.23.0

  • Added: a “Message box position on phones” option under Display Settings > Mobile Display, for when the chat fills the screen. On iPhone, the buttons that float above the keyboard could cover the message box. “Bottom, clear of the keyboard buttons” adds space under the message box while the keyboard is open, and “Top, under the header” moves the message box away from the keyboard. The default, “Bottom”, keeps the current layout. Thanks to @awmaeder for the report and screenshots.

1.22.1

  • Fixed: on small screens, a preset question that was typed into the message box wrapped onto a second line that could not be seen. Long questions now stay on one line and scroll to the left as they are typed; with reduced motion turned on, the question is cut off with “…”. Thanks to @awmaeder for the report.

1.22.0

  • Added: an option to type the preset questions into the empty message box, one after another, so visitors can see what they can ask before they start typing. It pauses while the visitor is typing, and visitors who have reduced motion turned on see the first question without animation. Turn it on under Chat Settings > Preset Question Buttons; it is off by default. Thanks to @awmaeder for the suggestion.

1.21.7

  • Fixed: embeddings did not reach 100% on their own. When the embedding provider’s rate limit was reached (common on the Gemini free tier), “Generate Embeddings” stopped and site learning skipped the chunks, with nothing retrying them later. “Generate Embeddings” now waits and continues, and anything still missing is filled in by a background job every few minutes until it is done. Thanks to @awmaeder for the report.
  • Fixed: the “remaining” count while generating embeddings showed at most 1. It now shows the real number.

1.21.6

  • Fixed: on sites whose options table still uses the older “utf8” character set, no chatbot settings could be saved, including API keys. The settings contain an emoji (the default bot avatar 🤖), and such a table cannot store emoji, so the database rejected every save without an error. Emoji and other characters the table cannot hold are now stored as character references and read back as the original characters. Sites on “utf8mb4” are unchanged. Thanks to @awmaeder for the report and testing.
  • Added: a notice on the plugin’s screens when the database rejects a settings save, instead of failing silently.

1.21.5

  • Fixed: after the setup box at the top of the Settings page connected a Gemini or OpenRouter key, clicking “Save Settings” without reloading switched the AI provider and model back to their old values, so the key looked lost. The settings form now updates itself as soon as the connection succeeds. Thanks to @awmaeder for the report.

1.21.4

  • Added: Settings > Security > “Load reCAPTCHA: Only when the chat is opened”. Google’s reCAPTCHA script (about 700 KB on mobile, which PageSpeed lists twice because reCAPTCHA loads it again in its own hidden frame) is then loaded when a visitor opens the chat or clicks into its input, instead of on every page load. Pages where nobody chats stay lighter. The first message, a lead form or an offline message waits for the script before sending. Off by default, so nothing changes until you switch it on. Thanks to @slafever for the report and measurements.

1.21.3

  • Fixed: conversation times ignored the WordPress timezone. The database server records conversation and message times in its own timezone, and the Conversations list, Conversation Details and the CSV export showed them without converting, so on hosts whose database runs in another timezone they were hours off. They are now shown in the timezone set in Settings > General, existing conversations included. The Knowledge list and the learned-pages list are corrected the same way. Thanks to @slafever for the report.
  • Fixed: the Conversations date filter and the “today” counts on the dashboard used the database server’s day instead of the site’s.

1.21.2

  • Added: during Black Friday 2026 (November 26 to December 2), the Settings screen shows a short notice about the Rapls AI Chatbot Pro discount, with the end time in your site’s timezone. It appears only on the Settings screen, only to administrators of sites without Pro, and not at all outside those dates. Each administrator can dismiss it for good, and the raplsaich_show_promo_notice filter turns it off. Nothing is loaded from outside the plugin.
  • Changed: while that notice is shown, the review request is held back, so the two never appear together. Once the notice is dismissed or the dates have passed, the review request returns under its usual conditions.
  • Added: a “Pro” link in this plugin’s row on the Plugins screen while Pro is not active. The raplsaich_show_pro_row_link filter removes it.

1.21.1

  • Fixed: in the free version, Site Learning could not be turned back on once it was off. The switch was only on the Site Learning screen, which appears when the Pro add-on is installed, and the dashboard’s “Run site learning” link led to the Knowledge page. Settings > AI Settings now has a Site Learning switch, the number of learned pages and the last run, and a “Run Learning Now” button. Choosing content types and the schedule, and managing learned pages, stay in Pro. Thanks to @solroed for the report.
  • Fixed: when Site Learning was on but its daily schedule had gone missing, learning stopped without any sign. The schedule is now restored the next time wp-admin is opened.
  • Changed: the readme now says which parts of Site Learning and conversation history are in the free version and which are in Pro, and describes the retention setting correctly (a number of days; 0 keeps conversations indefinitely).

1.21.0

  • Fixed: Gemini 2.0 Flash, the default Gemini model, was shut down by Google on June 1, 2026, and Gemini 2.0 Flash-Lite and Gemini 3 Pro Preview have been shut down as well. A site still set to one of them now sends its chats to Google’s replacement (Gemini 3.6 Flash, Gemini 3.1 Flash-Lite and Gemini 3.1 Pro Preview), and wp-admin says so until a model is chosen again. New installs use Gemini 3.5 Flash-Lite. The saved setting is not changed for you.
  • Fixed: OpenAI shut down o3-mini on October 1, 2026, and shuts down GPT-4.1 nano and o4-mini on October 23 and GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 Pro and o3 on December 11. From each date, a site still set to one of these models sends its chats to OpenAI’s replacement (GPT-5.6 Sol, Terra or Luna). Claude Sonnet 4.5, which Anthropic retires on November 30, 2026, moves to Claude Sonnet 5.5 the same way.
  • Added: a warning, on the settings screen and on the plugin’s own admin screens, when the selected model has an announced retirement date. It says when, and which model will answer afterwards.
  • Added: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and GPT-5.6 Sol, Terra and Luna; Claude Sonnet 5.5 (now the recommended Sonnet) and Claude Opus 5.5; Gemini 3.8 Flash (now recommended), 3.7 Flash, 3.6 Flash, 3.5 Flash, 3.5 Flash-Lite and 3.1 Flash-Lite. Models that are shut down or scheduled to be are removed from the lists; a site still using one keeps it selected, and a save no longer switches it to the first model in the list.
  • Changed: retired Claude Sonnet and Opus models now go to Claude Sonnet 5.5 and Claude Opus 5.5 instead of the 4.6 models.
  • Changed: GPT-6 and Gemini 3 models are sent without temperature (GPT-6 rejects it, and Google advises against lowering it on Gemini 3) and get more room for thinking, so replies are not cut off. Claude models that think by default, such as Sonnet 5.5 and Opus 5.5, are asked for low effort.
  • Fixed: with web search on and nothing found in the site’s content, Claude Opus 5.5 and Claude Fable 5.1 returned an error, because the plugin forced the search tool, which those models do not accept. They are now told to search in the prompt instead.
  • Fixed: the OpenAI-compatible list offered deepseek-chat, which DeepSeek discontinued on July 24, 2026, and glm-4-plus, which no longer answers; the OpenRouter list offered deepseek/deepseek-chat-v3, which does not exist. They are replaced by deepseek-flash, glm-5.3, qwen3.8-max and deepseek/deepseek-chat, and the OpenRouter Claude entry is now Claude Sonnet 4.6. The model field’s help text no longer names DeepSeek or GLM models, which change too often.
  • Fixed: usage costs. GPT-4.1 models were counted at the GPT-4 price, up to 150 times too high, which could stop the chat early under Pro’s monthly budget limit, and most current OpenAI, Claude and Gemini models had no price, so GPT-5.2 Pro, for example, was counted at about 1/280 of its cost. Prices now follow each provider’s pricing page.
  • Changed: Model Fallback uses Gemini 3.1 Flash-Lite, and the free-tier setup picks a Gemini 3 Flash-Lite model, because Google now serves Gemini 2.5 only to projects that already used it.

Older versions

The changes in 1.20.5 and earlier are listed in the 1.20.5 readme.