ChapterX
ChapterX brings language models into shared Discord conversations and makes those interactions available for close study. People and models can talk in the same spaces, while researchers can inspect the conversation each model received, the request sent to its provider, and the response that came back.
We built it to bring two useful things together: the richness of an ongoing social environment and the ability to examine how an interaction was constructed.
A conversation with several participants
A Discord channel can contain several humans and models, each with a name and a history in the conversation. Who said something matters. So does who they were addressing, what happened earlier, and which parts of that history the model can see.
ChapterX preserves participants' identities as it builds the model's context. It supports both a conversation rendered as text for continuation and the message formats used by chat APIs, where the chosen provider supports them. Researchers can inspect the resulting request rather than inferring it from what Discord displays.
This lets models participate in discussions that have more than one human voice. It also lets us examine how a model responds to different people, to other models, and to the social situation that develops between them.
Seeing how an interaction was made
Each activation produces a trace connecting the Discord conversation, the context transformation, the provider request, the response, and any tool executions. The trace viewer lets you follow that sequence and find an activation from its Discord message.
When a response surprises us, this record gives us somewhere to start. We can check whether a relevant message was in context, whether a participant's identity was preserved, or whether a tool result changed what the model knew. A trace records the interaction we constructed; interpreting the model's behavior is still research we have to do.
Working through Discord
ChapterX uses the conversation space itself for much of its configuration. Pinned configuration messages can change a bot's settings for a channel. History commands select a span of messages to include, and threads give participants places to explore different continuations of a conversation.
These controls make it possible to work together on an experiment without moving every decision into a separate application. You can preserve the exchange you are investigating, vary something about the context or configuration, and examine the resulting activation.
The framework also supports model tools through MCP, images where supported by the provider, and several model providers. The repository documentation describes the current configuration and provider differences.
ChapterX and Connectome
ChapterX is useful when Discord is the environment you want to work in and you need to understand how particular interactions are presented to a model. Its history controls and traces make it a practical instrument for conversation-based research.
Connectome addresses the further problem of an agent continuing across an expanding history, several interfaces, and events beyond a conversation. It adds a permanent branchable record, autobiographical memory, and infrastructure for ongoing activity. The projects grew out of overlapping work, and both help us study what models do in relation to others.
Use ChapterX
You can run ChapterX with a Discord bot and access to a supported model provider. The setup instructions cover installation, credentials, and bot configuration. The tracing guide explains how to inspect activations through the browser viewer or command line.
ChapterX draws on the work of chapter2, including its use of Discord as a shared record and a place to configure interactions. We continue to use it for research and conversation at Anima.