Plugins
AI plugins for Bubble
Plugins that add AI agents, document search, data analysis and live voice to Bubble apps, sold to the developers who build on it.
About them
Bubble is a no-code platform, and plugins are how its developers add the things it can't do on its own. These are my AI and real-time ones. Each is written in JavaScript, sold on Bubble's marketplace, and supported by me when something goes wrong in someone's live app.
They're where I learned most of what I know about making AI dependable: keeping API keys off the page, capping how many steps an agent can take, retrying when a tool call fails, and handing results back in a form a developer can use straight away without unpicking JSON.
Agents and data
A voice and text agent for any Bubble page. It runs on Groq or Gemini, listens through Deepgram, speaks through Cartesia, ElevenLabs or OpenAI, can see through a webcam, and answers from your own documents with RAG, optionally backed by Pinecone. Function calls can be intercepted before they're sent on. It's the main plugin behind Interactive Avatar.
An agent that takes action rather than just chatting. It finds your Backend Workflows by a naming convention and uses them as tools, running the whole loop of deciding, calling and reading the result inside one action. It works with OpenAI, Anthropic or Gemini, and you can switch without touching a workflow. There's a cap on how many steps it can take, automatic retries when a tool fails, token counts for every turn, and a server-side version that runs from a schedule or webhook with no page open.
Questions and answers over your own content: text, PDFs, and images, which a vision model describes so they can be searched too. Everything is stored and searched inside the app's own Bubble database, with no outside vector store. Answers come back with their sources, follow-up questions are rewritten before the search so the context carries over, and each user's data is kept separate. It only uses OpenAI, on purpose, because a RAG answer depends far more on what gets retrieved than on which model writes it.
Points AI at an app's database and hands back structured results. There are seven built-in analyses, from anomaly detection and sentiment to spotting missing or contradictory data, plus plain-English questions answered with the IDs of the records the answer came from. It can follow relationships up to three levels deep and write generated content straight back to each record.
Access to hundreds of models hosted on Hugging Face, either through their inference API or run directly in the browser using a worker file, across more than 25 different tasks.
Voice, video and avatars
A photo-realistic avatar that visitors talk to out loud, powered by the OpenAI Realtime API. Session tokens are created on the server so no API key ever reaches the browser, the spoken language is fixed from the first word, and a per-visitor limit caps how many sessions someone can start.
Video and audio calling on LiveKit, built up to match Google Meet and WhatsApp: screen sharing, background blur, end-to-end encryption, and avatars that follow your face in place of your camera. Audio can also be sent over a WebSocket to feed other services.
An audio recorder that works in the browser, with live transcription from either the browser's own speech recognition or OpenAI's realtime service, and a one-step upload to Cartesia for voice cloning.
Animates an element while someone talks, combining the microphone level with speech recognition, so it can tell when someone is speaking with their microphone muted.
OpenAI's realtime speech-to-speech over WebRTC, kept deliberately simple.