[opensourcedex]

thoughtdag is a free, open source note taking & knowledge management project written in TypeScript and released under MIT. It has 511 GitHub stars, 50 forks and 5 open issues, and was last pushed 6 hours ago. On this registry it ranks #56 of 60 tracked projects in Note Taking & Knowledge Management, with 5 head-to-head comparisons available.

What is thoughtdag?

ThoughtDAG is an MIT-licensed TypeScript application that turns LLM conversations into an editable thought graph on an infinite canvas, where each exchange becomes a node and the wires between nodes, rather than the whole transcript, decide what the model sees next; it is built for people who branch into side questions and want to keep the useful paths.

What it is

ThoughtDAG presents an LLM conversation as a graph rather than a single thread. Every exchange becomes a node on an infinite canvas, and a user can branch from a detail, explore it separately, and then connect the paths that should inform a later question. The rule the project states is that "wires are the context": connecting conversation paths adds them to the next request, and disconnecting a path removes it from the context without deleting the work. The graph changes the model's input, not merely the layout, and the desktop app also lets a document be read beside the conversation so a branch can be taken from a passage.

The problem it solves is context selection in chat. In an ordinary interface the context is the accumulated transcript, so earlier work is recovered by replaying a whole session. ThoughtDAG instead keeps a local index of supported agent sessions and ThoughtDAG canvases, and its topic dossiers collect decisions and open questions with links back to their sources. It lives in the TypeScript and Node.js ecosystem, distributed through npm, a Homebrew cask and plugins for DeepSeek Harness, and it is local-first, with the user supplying the model connection. The repository has 504 stars, 50 forks and 5 open issues.

Key capabilities

  • An infinite canvas in which every exchange is a node and the wires between nodes determine context; the project's one rule is that wires are the context.
  • A preview of what the model will receive before a request is sent, with wires selecting the conversation paths that are included.
  • Command-line search over the local index: npx thoughtdag why src/lib/api.ts finds conversations about a file, npx thoughtdag find "a phrase you remember" finds matching turns, and npx thoughtdag topics lists indexed topics.
  • Read-only history tools for agents through thoughtdag setup mcp, so relevant turns are retrieved rather than a whole session replayed.
  • An optional Jev decision layer that identifies topics and ranks relevant excerpts; without a decision model, recall falls back to rules, and the context panel lists the dossiers and excerpts added to a request so they can be inspected or excluded.
  • A DeepSeek Harness plugin, dsh-thoughtdag, that bundles the canvas and memory layer and switches between chat and the graph inside the harness.

Who uses it and how

  • Desktop users on macOS, Windows or Linux who install the cask or a download, connect their own model and open the example canvas, reading a document beside the conversation and branching from a passage.
  • DeepSeek Harness users who add the plugin and choose context on the canvas while the harness runs the next turn, on Node 22.19 or later within the 22.x line or Node 24 or later, with DeepSeek Harness 0.1.2-rc.1 or later.
  • Developers and writers who remember a file, a phrase or a URL but not the session, and search local conversations from the terminal to jump to the matching turn without opening the desktop app.
  • Agent builders who expose read-only history tools with thoughtdag setup mcp.

Getting started

Install the desktop app with brew install --cask thoughtdag on macOS, or download a build for macOS, Windows or Linux, then connect a model and open the example canvas. For command-line use, run npx thoughtdag directly or npm install -g thoughtdag for regular use; dsh plugin --profile web add dsh-thoughtdag followed by dsh web installs the DeepSeek Harness plugin.

How it compares

The facts supplied name no comparable or paid products for ThoughtDAG to be measured against, so it stands alone in this registry rather than being positioned against named alternatives. The distinction the project draws for itself is that the graph changes the model's input rather than the layout of the conversation.

When to use it — and when not to

A user must supply their own model connection and maintain a local index, and ranked recall depends on the optional Jev decision layer, since recall falls back to rules without a decision model. Anyone who wants a hosted, zero-setup chat product, or who cannot meet the runtime requirements — Node 22.19 or later on the 22.x line or Node 24 or later, and DeepSeek Harness 0.1.2-rc.1 or later for the plugin — should not choose it. The published timing comparison is narrow by its own terms: selection-stage latency only, six runs per engine over 14 synthetic excerpts, excluding retrieval and answer generation.

project readme (upstream, from github) — read inline

ThoughtDAG

AI conversations that branch on an infinite canvas.

Each exchange becomes a node. Wires are the context.
Explore a side question, connect useful paths, and choose what the model sees next.

Download · Website · Docs · 中文

License

0.5 update · CLI · Harness · Desktop · How it works · How it differs · Research

New in 0.5 · ThoughtDAG × Jev

Bring relevant past conversations into the question you are asking now.

  • Find earlier work. The local index searches supported agent sessions and ThoughtDAG canvases. Topic dossiers collect decisions and open questions with links back to their sources.
  • Select what belongs. The optional Jev decision layer helps identify topics and rank relevant excerpts. Your chosen language model develops the answer.
  • Check what comes back. With recall enabled, the context panel lists the dossiers and excerpts added to a request. Inspect their sources or exclude individual items before continuing.

In a small relevance-selection pilot, Jev's median was 391 ms versus 24,813 ms for our GLM adapter. These are selection-stage timings, not end-to-end search or answer times.

What the timing measures

Six runs per engine over the same 14 synthetic excerpts. Median selection latency: 391 ms for Jev-1.13 and 24,813 ms for the GLM-5.3-Flash adapter with default reasoning. These are different inference paths, not a controlled ranking of model speed. Retrieval and answer generation are excluded; this does not measure whole-product speed or accuracy gains.

Without a decision model, recall falls back to rules. The System 1 / System 2-style split describes software roles here: quick relevance decisions, then answer and dossier generation. It is not a claim about human cognition.

Set up history and recall · Configure Jev

Find past context from the command line

Remember a file, a phrase or a URL, but not the session? Search local conversations and jump to the matching turn, without opening the desktop app.

npx thoughtdag why src/lib/api.ts           # conversations about this file
npx thoughtdag find "a phrase you remember" # matching conversation turns
npx thoughtdag topics                       # topics in your local index

For regular use: npm install -g thoughtdag. Run thoughtdag setup mcp to expose read-only history tools to your agent. Retrieve the relevant turns rather than replaying a whole session. CLI guide →

Inside DeepSeek Harness

Switch between chat and ThoughtDAG's graph inside the harness. Choose the context on the canvas; the harness runs the next turn.

dsh plugin --profile web add dsh-thoughtdag
dsh web

The plugin bundles the canvas and memory layer. Requires Node 22.19+ (22.x) or 24+, and DeepSeek Harness 0.1.2-rc.1 or later. Plugin guide →

The desktop app

Read a document beside your conversation, branch from a passage, and connect the paths you want to explore together. Use your own model connection.

brew install --cask thoughtdag

Or download for macOS, Windows or Linux, connect a model and open the example canvas.

Official YouTube thumbnail: ThoughtDAG narrated tour

▶ Watch the 33-second tour

The one rule

Wires are the context. Connect conversation paths to use them in the next question. Disconnect a path without deleting the work.

Branch from a detail, explore it separately, then connect the useful parts to a later question. The graph changes the model's input, not just the layout.

Preview what the model will receive before sending. Wires select the conversation paths; explicit references and enabled recall can add material alongside them. Context guide →

In action

✂️ Change the context, keep the exploration

Select text in an answer to start a side branch. Disconnect that branch from a later question, then regenerate to compare. Its nodes stay on the canvas: keep exploring from them or reconnect them later.

📖 Read, clip, and ask

Open a PDF, image or HTML alongside the graph. Ask about a passage or clip a figure into its own node. PDF clips keep their page reference, so you can check the source as the discussion develops.

💎 Condense the path; weave the highlights

Condense creates a shorter copy of a conversation path while preserving the original. Weave turns selected highlights into cited prose. Continue from the result, or export it as Markdown. Zooming out changes the view, not the context.

🧭 Session Atlas: continue an earlier conversation

Open a supported local agent session as a graph. Pick where to branch or continue; use the history index to find related discussions from other sessions. Atlas provides the view, and recall helps find what to bring in.

Supports local Claude Code, Codex, DeepSeek Harness and Pi sessions. Source sessions remain read-only.

How ThoughtDAG differs

Nodes and edges serve different purposes. Here is where ThoughtDAG fits:

Product category ThoughtDAG's focus
Linear chat Keep several lines of inquiry visible and choose which ones continue into the next question.
Mind maps and whiteboards Use connections to change model input, not just organize ideas visually.
Branching chat canvases Connect several branches into one question, or disconnect a path while keeping its nodes.
Agent workflow canvases Edit conversational context as you explore, rather than design a pipeline of automated tasks.
Retrieval and automatic memory Inspect source-linked dossiers and recalled excerpts; edit or exclude what the next request uses.
Code graphs and conversation search Find the discussions behind a file or topic across supported agents, then continue from them.
Harness context viewers Move from inspecting a session to composing and sending its next turn.

These categories overlap; individual tools may share capabilities. ThoughtDAG is not an autonomous research agent or a replacement for your coding harness. Retrieval can miss relevant history, and generated dossiers still need checking.

🗺️ Export the shape of your thinking

Export the canvas as a Thought Map: nodes, wires and structural counts, without the full conversation text. Use it to share how an investigation branched, narrowed and came together.

More ways to run

Run from source

npm install
npm run server    # LLM proxy :3001
npm run dev       # frontend :5173

Configure a model in the app or through environment variables. Local setup →

Browser demo

The browser demo includes an example canvas that needs no API key. It is a subset: local session discovery, Session Atlas and the local history/memory layer require desktop or local hosting.

🧪 Research: Why editable context matters

Context Intervention Benchmark · Pilot v2

9 model endpoints · 1,485 scored responses · exact-match scoring

Deleting a wrong claim may leave its consequences in later replies. In our synthetic pilot, removing the source alone repaired 152 of 162 affected model-cases; removing the contaminated subgraph repaired 162, and recomputing descendants repaired 161. The report includes the protocol, results and limitations. This is a context-intervention experiment, not a general model leaderboard.

Read the case study · Methods and results · Suggest a model

More capabilities

Capability What it adds to the same workflow
Request preview Check the conversation, references and recalled material assembled for the next call.
Staleness and replay Review dependent answers after an upstream edit; rerun in dependency order.
Per-node model selection Try a different model on a branch without changing the entire canvas.
Read-only sharing Share a graph for others to inspect; preview its contents before publishing.
Folder backup Save canvases as local files and keep a recoverable copy outside browser storage.

Full capabilities and roadmap →

Models, cost & privacy

Canvases, documents, the index and dossiers are stored locally. Remote model calls send relevant content to your configured providers, including decision and dossier-generation calls. Provider charges may apply; disabling Jev does not disable ordinary model calls.

Connect local Ollama or an OpenAI-compatible endpoint. Inside DeepSeek Harness, inference uses the harness's providers and keys. Export backups and Markdown, and review text and metadata before sharing. Setup and privacy details →

Contributors

@KehanLiu @nasodaengineer @hexu321 @Moya-Doc @nanami-0713 @LHN-xiao-hai-tun @HarveyZed @Pireirik

Contributions are welcome — start with CONTRIBUTING.md.

Supporters

With gratitude to @andreilaiter, ThoughtDAG's first supporter, and to everyone helping this independent open-source project grow.

MIT © 2026 Xia Chen · Roadmap · Feedback · Cite

Frequently asked questions

Is thoughtdag free to use?

thoughtdag is open source under the MIT licence. There is no licence fee and no seat count — you can self-host it or, where the project offers one, pay a vendor for a managed version instead.

What does thoughtdag do?

Your thinking deserves a map: an infinite canvas where LLM conversations grow into an editable thought graph. Wires are the context.

What is thoughtdag written in?

thoughtdag is primarily written in TypeScript. Its source is publicly available at https://github.com/chenxiachan/thoughtdag, and it has 511 GitHub stars.