FigJam now hosts Jambot, a widget that puts OpenAI's ChatGPT inside a multiplayer canvas so teams can generate, condense and remix ideas without leaving the board. Daniel Mejia, Sam Dixon and Aosheng Ran built it at an internal AI hackathon and carried it from demo to shipped widget in a few weeks, with help from other Figma collaborators.

Why the chat box wasn't enough

Mejia pitched the project as "a visual version of ChatGPT," arguing that while the plain chat interface is powerful, it isn't the optimal format for every scenario. When the model returns multiple options, following up on one means scrolling back and re-asking questions to revisit another; a one-dimensional message sequence also makes it awkward to branch into different topics or tweak ideas to see how they connect.

His reference points were network-of-thought tools such as Roam Research and Logseq, which let pages link to one another so ideas can be connected, organized and traced, plus Albus, which gives AI interaction a visual feel. Dixon brought a complementary view from LangChain — object-oriented and semantic, but requiring code — and wanted something visually tangible and easy to use.

Ran framed the problem more broadly: "I think we're stuck in chat boxes. Just like we're stuck in Zoom right now." In his view, the GUI primitives built over the past 20 years to make computers approachable have yet to be reinvented for AI, leaving users with something closer to rudimentary command-line models like DOS.

Functions instead of a blank prompt

The team ruled out shipping just a text box that submits an API request. A pre-set list of functions, Dixon said, gives users clear affordances and makes the capabilities discoverable without requiring them to learn a new way of interfacing with a computer. Three functions emerged from those discussions:

  • Ideate — one sticky becomes many.
  • Summarize — many stickies become one.
  • Play — expand on ideas and summaries; one-to-one transformations such as turning feedback into a song, which the team treats as the tool's opportunity for whimsy.

Because GPT models "tend to be really good at doing," Mejia said, folding those prompts into Jambot felt natural.

The canvas as the interface

The design settled into an almost game-like, low-stakes form: two node connectors, one at an input space and one at an output space, with information flowing left to right — extracting content from a sticky or running a command, then emitting a result. Ran describes it as applying the visual programming interface to ChatGPT, a combination that "just feels right as a tool."

Running on FigJam gives Jambot visual and multiplayer capabilities out of the box, which Mejia sees as its differentiator from other AI products and a natural fit for collaborative work. Function suggestions surfaced automatically as next steps also keep people in a flow state, pointing them toward uses of AI they wouldn't have thought to type into an empty box — "why don't you turn this into a poem," for example.

Assistant, not oracle

Figma's stated vision is AI as a tool to supercharge work and collaboration — less about doing things for you, more about helping you do a better job. Dixon ties Jambot to that framing: it lets you play with an idea or piece of information rather than complete a clearly defined task, which positions AI as an assistant rather than "an oracle of some sort" that hands down a single answer with no way to intervene.

The output persists as a visual artifact capturing the thinking and learning process, which can be revisited or shared later. Mejia's ambitions go past text in and text out — he is eager to see how Jambot helps people dive into different paths fluidly and open up ideas that can be visualized and tinkered with. The team is soliciting feedback to inform which features to pursue natively.