➡️➡️➡️ The Complete AI Toolkit for Bubble 🤖✨
Chat, voice, vision, documents, knowledge bases and structured data in a single plugin. A finished streaming chat you drop on a page, a voice assistant that holds a real spoken conversation, and 21 server actions covering everything from OCR to image generation. Every piece shares the same fields across OpenAI, Claude and Gemini, so changing provider is a dropdown.

✅ 5 visual elements — chat, voice, streaming text, autocomplete and a ⌘K palette
✅ 21 server actions — vision, PDFs, images, speech, embeddings, web search, moderation
✅ 3 providers behind one set of fields, states and events
✅ Light & dark themes, automatic or forced, with 12 appearance fields on the chat
✅ Live voice over WebRTC without your API key ever reaching the browser
✅ Structured output validated against the shape you asked for
✅ Knowledge bases that live in your own Bubble database
✅ 580 automated tests across transport, streaming, schemas and every error path

────────────────────────────────
CHECK IT
────────────────────────────────

Live demo: https://plugins-testing-calendar.bubbleapps.io/version-test

Editor (see how it is set up):https://bubble.io/page?id=plugins-testing-calendar&tab=Design&name=index

────────────────────────────────
THE FIVE ELEMENTS
────────────────────────────────

💬 AI Chat — a chat surface that is already finished
Answers stream in word by word. Markdown renders as it arrives: bold, lists, links, code blocks. Suggested prompts, a Stop button, Regenerate, and conversation history you can save to your database and restore later. Nine colours, two corner radii and a font, or leave it and it follows the visitor's light/dark setting. Fill in the tools field and the same element becomes an agent that runs your Bubble workflows mid-answer.

🎙️ AI Voice — it answers the phone
A spoken conversation over WebRTC. The visitor talks, it talks back, and they can cut it off mid-sentence. Live transcription of both sides. It runs your workflows during the call: check availability, write a booking, look up an order, and say the result out loud. Your private key stays on the server the whole time, because the browser only ever gets a token that expires in about a minute.

✍️ AI Text — the engine, none of the interface
Streaming text inside your own design. The response state grows as it arrives, so you can put it wherever you like: a product description, a rewritten email, a summary. It runs in the browser, which means Bubble's 30-second server ceiling does not apply.

⌨️ AI Autocomplete — it writes with your data
Grey text suggesting how to continue, accepted with Tab or a tap. Point its context field at the row somebody is working on and the suggestions use those actual numbers: the square metres, the plan name, the delivery date. Not generic filler.

⌘ AI Command — type it instead of clicking it
A ⌘K palette over your own workflows. Someone types "give Nova Solutions 20% off their next order" and you get back a command name and its arguments, ready to run.

────────────────────────────────
THE 21 SERVER ACTIONS
────────────────────────────────

Your private key never leaves the server.

📝 Text
✅ Generate Text — one question, one answer, with validated structured output if you want it
✅ Transform Text — 15 operations in one dropdown: summarise, rewrite, improve, expand, shorten, translate, change tone, simplify, proofread, classify, keywords, entities, sentiment, detect language, SEO
✅ Universal AI Request — any provider parameter the plugin does not model, plus the raw response
✅ Get JSON Value — read any field out of a JSON string, which Bubble cannot do on its own

🖼️ Images and documents
✅ Analyze Image — describe, alt text, OCR, extract fields, classify, list objects, compare several
✅ Analyze Document — from a PDF: summarise, ask it questions, extract fields, analyse, classify, translate, pull the full text
✅ Generate Image — create from a description, three shapes, transparent background
✅ Edit Image — change an existing image, with a mask to touch only part of it

🔊 Voice
✅ Transcribe Audio — audio to text, with language and your own vocabulary
✅ Text to Speech — text to audio, 11 voices, speed and delivery style
✅ Create Voice Session — mints the short-lived token for AI Voice

📚 Knowledge
✅ Create Embedding — text to a meaning vector, one or many at once
✅ Compare Texts — how alike two texts are by meaning rather than spelling
✅ Index Text — split a document and prepare it for search
✅ Search Knowledge — find the passages that answer a question
✅ Ask Knowledge — answer citing the sources, with or without an index
✅ Search Web — search the live web and answer with links

🧠 Memory
✅ Remember — store something about a person for later
✅ Recall — bring back what is relevant, for that person only

🛡️ Safety and data
✅ Moderate — is this text or image safe? With category and confidence
✅ Analyze Data — over a list: classify, rank, clean, spot anomalies, extract fields, recommend, summarise, deduplicate

────────────────────────────────
OTHER FEATURES
────────────────────────────────

🎯 Structured data without writing JSON Schema
Describe the shape in plain text and the answer comes back checked.

📚 A knowledge base inside your own Bubble database
The embeddings are quantised so a passage costs about a third of a kilobyte, which means the index fits in an ordinary Bubble field. No vector database to sign up for. Under roughly 30 pages you do not even need an index, and the plugin says so rather than selling you the complicated version. Above 500 pages, Pinecone is supported.

🌐 Web search with no second API key
Today's rates, live prices, whether a flight is delayed. All three providers have search built in, and this uses it with the key you already have.

🔑 Bring your own key
Every action and element takes an optional api_key field. Leave it empty and your key is used. Fill it and each customer, or each visitor to a demo page, pays for their own calls. Multi-tenant AI without running a proxy server.

🛠️ Agents that run your Bubble workflows
Declare tools in plain text. The assistant decides when one is needed, your workflow runs, you hand the result back, and it carries on with the answer. Step limits and timeouts are built in, because an autonomous loop that bills per iteration is a hazard. A tool the model invents gets refused instead of run.

💰 Straight about cost
Every answer reports its tokens and an estimated cost. Where a model's price cannot be verified the cost field is left blank rather than showing a confident $0.00. You can put that number on screen for your own users.


────────────────────────────────
ONE EXAMPLE YOU COULD BUILD
────────────────────────────────

A customer rings your repair shop. Nobody is at the desk.

🎙️ AI Voice picks up and says hello. They explain that the frame on their bike has cracked. Mid-sentence the assistant runs one of your Bubble workflows, a tool you declared in a text box, which calls Ask Knowledge over your warranty documents and comes back with "Fjord 3 frame, covered five years, this is inside it".

Still on the call, it runs a second workflow that searches your database for the nearest workshop with a free slot, and says out loud: "That’s covered. Gràcia has an opening Thursday at 10 AM, just two kilometres away. Would you like me to book it?" They say yes, and the booking is written to your database before they hang up.

The call ends. The transcript of both sides is saved to the ticket, Transform Text turns it into a one-line summary for your team, and the customer gets a confirmation.

No phone system. No third-party call centre. No proxy server. Five pieces of one plugin, wired together with Bubble workflows.

────────────────────────────────

Full documentation with a worked examples: https://app.notion.com/p/AI-Toolkit-Documentation-3c761f4cca3a80fbbf13fbaaf6aa7e3d?source=copy_link