← all repositories
chenxiachan/thoughtdag

A chat canvas where the wires are the prompt

ThoughtDAG turns LLM conversations into an editable graph, so you choose exactly which branches feed the next question instead of hauling the whole scrollback along.

thoughtdag
Velocity · 7d
+2.6
★ / day
Trend
↘cooling
star history

What it does Every exchange becomes a node on an infinite canvas. Branch off a side question, explore it, then wire the useful paths into a later question — or disconnect a dead end without deleting it. The connections aren’t decoration: they decide what the model actually receives, and a request preview shows the assembled context before you send it. PDFs, images and HTML sit alongside the graph, and clipped passages keep their page references.

The interesting bit In most canvas chat tools, the graph is interior decorating — nodes you drag around. Here the edges are load-bearing: rewiring the canvas literally rewrites the model’s input. The authors even ran a synthetic context-intervention pilot (9 model endpoints, 1,485 scored responses) to argue the point: when a wrong claim contaminated downstream answers, deleting just the source repaired 152 of 162 affected cases, while removing the contaminated subgraph repaired all 162.

Key highlights

  • Branch from any passage; disconnecting a path drops it from context while its nodes stay on the canvas for later.
  • Request preview shows the conversation paths, references and recalled excerpts the model will actually receive.
  • Session Atlas opens local Claude Code, Codex, DeepSeek Harness and Pi sessions as graphs — sources stay read-only — so you can continue an old thread from any branch point.
  • After an upstream edit, dependent answers can be reviewed for staleness and rerun in dependency order; each node can use a different model.
  • Desktop app with local-first storage, Ollama or any OpenAI-compatible endpoint; a CLI and read-only MCP tools let agents pull past turns instead of replaying whole sessions.

Caveats

  • Local-first covers storage, not inference: remote model calls send relevant content to your configured providers, and disabling the Jev decision layer doesn’t disable ordinary model calls.
  • The README admits retrieval can miss relevant history and generated dossiers still need checking — it’s explicitly not an autonomous research agent or a coding-harness replacement.
  • The flashy Jev figure (391 ms vs 24,813 ms) is a selection-stage median over 14 synthetic excerpts, not end-to-end search or answer speed.

Verdict Worth a look if your LLM sessions branch, loop back, and drown in their own scrollback — doubly so with a local model. If you wanted an autonomous agent or a harness replacement, the README is refreshingly upfront that this isn’t that.

Frequently asked

What is chenxiachan/thoughtdag?
ThoughtDAG turns LLM conversations into an editable graph, so you choose exactly which branches feed the next question instead of hauling the whole scrollback along.
Is thoughtdag open source?
Yes — chenxiachan/thoughtdag is open source, released under the MIT license.
What language is thoughtdag written in?
chenxiachan/thoughtdag is primarily written in TypeScript.
How popular is thoughtdag?
chenxiachan/thoughtdag has 800 stars on GitHub and is currently cooling off.
Where can I find thoughtdag?
chenxiachan/thoughtdag is on GitHub at https://github.com/chenxiachan/thoughtdag.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.