MulmoChat adds live visual canvases

On May 20, open-source project MulmoChat surfaced in an X post showing a chat interface that turns model output into a live canvas, including map- and mind map-style views, rather than leaving everything in a text thread. The project’s GitHub repository describes MulmoChat as a “research prototype” for multimodal AI chat in which “rich visual and interactive content” appears directly on canvas during the conversation. (github.com) What makes the demo notable is the model layer. MulmoChat’s roadmap and server design documents indicate it is built to route generation through multiple providers behind a shared contract, including OpenAI, Anthropic and Google Gemini, with the goal of switching model backends without breaking the canvas or tool flow. A planning document says the server now exposes consolidated text-generation endpoints that proxy OpenAI, Anthropic, Google Gemini and Ollama. (github.com) The broader idea is that the chat window is no longer the whole product. In MulmoChat’s own documentation, the project is presented as an attempt to merge natural-language interaction with graphical interfaces so that “images materialize, maps become explorable, games turn playable” inside the same conversational flow. That puts it in the same interface conversation as artifact-style and canvas-style AI products, but MulmoChat is positioning itself as an open-source prototype rather than a closed commercial feature. (github.com) The repository also suggests this is still an active build, not a finished release. GitHub showed about 170 stars and roughly 40 forks when the page was crawled, and the latest listed commit was last week. The same repo includes planning documents for provider switching and a separate “MulmoClaude” concept, which describes a next-generation agent app built on Claude Code and the project’s GUI chat protocol. (github.com) One caveat: some of the most specific claims in the social post — especially collaborative editing and export formats such as PNG, SVG and PDF — were not clearly spelled out in the repository excerpts available through search results. The GitHub materials do support the core claims that MulmoChat is open source, canvas-based, multimodal and designed around multiple model providers, but the export and collaboration details appear to come primarily from the May 20 demo post rather than the repo text surfaced in search. (github.com) So the clean read is this: MulmoChat is an open-source experiment in replacing linear AI chat with a live visual workspace, and its architecture is being built to work across OpenAI, Claude and Gemini rather than being tied to one model vendor. The next place to watch is the GitHub repo and any follow-up posts from the project’s maintainer, because that is where feature claims such as collaboration, export and backend switching are most likely to be documented in fuller detail. (github.com)