Linear AI chats hide their most important state: context.

ThoughtDAG makes that context visible as an editable directed graph. Each node is one question-and-answer exchange, while incoming wires determine exactly which earlier branches the model receives.

Branch an idea without damaging the original conversation. Merge selected conclusions into a new synthesis. Delete an irrelevant edge and regenerate the same prompt with cleaner context.

ThoughtDAG is open source, local-first, and human-controlled. It supports local Ollama models, arbitrary compatible endpoints, PDF reading with page-linked citations, local file backups, and JSON or Markdown export.

No account is required, and the included example canvas works without an API key.

I built ThoughtDAG because I kept losing control of context in long AI conversations. Branching was easy, but deciding which earlier ideas should influence a later answer was mostly invisible.

The experiment behind ThoughtDAG is simple: what if context were something you could inspect and edit directly, rather than something hidden inside a transcript?

I would especially value feedback on whether “wires are the context” feels immediately understandable after trying the example canvas.