How to Install ThoughtDAG on Windows, macOS or Linux

ThoughtDAG turns an AI conversation into an editable graph. Every wire determines which earlier questions, answers and materials the model receives next. Delete a noisy branch and it stops entering the prompt; join two useful branches and the next answer can use both.

To install ThoughtDAG, download the official desktop build for Windows, macOS or Linux, verify its checksum, and connect Ollama or a hosted API. The current v0.3.18 release has ready-made installers for all three platforms. This guide covers the desktop route and starts with a small graph before touching PDFs or coding-agent handoffs.

Quick answer: the recommended ThoughtDAG setup

1. Download the v0.3.18 installer from the official GitHub release.
2. Verify its SHA-256 hash before opening it.
3. Launch the included example canvas before connecting a model.
4. Use Ollama when local processing matters, or enter a real API key for a supported hosted provider.
5. Create two small branches, wire only the useful one into a new question, and inspect the context preview.
6. Turn on folder backup before building a large research map.

My recommendation: start with Ollama if the material is private and your computer can run a suitable model. A hosted API is usually the simpler route and may produce stronger answers, but prompts, selected context and extracted document text then go to that provider.

What ThoughtDAG does differently

A normal chat keeps growing, while a mind map may show relationships without changing what the model sees. ThoughtDAG makes the graph operational: incoming edges determine the next node’s context. You can retain a dead-end branch for your record without sending it again, or use a reference edge when only a selected result is needed.

The official feature reference describes several tools built around that idea:

• Context preview: see estimated tokens, messages and attached files before asking.
• Branching and merging: explore alternatives, then synthesize selected nodes.
• Staleness tracking: editing an upstream answer marks affected downstream nodes so they can be reviewed or replayed.
• PDF reading: select a passage, ask about it and retain a page reference.
• Local file backup: save the canvas as a real .thoughtdag.json file and export selected chains as Markdown.

It suits branching research and technical decisions. For a one-question chat, the graph is unnecessary overhead.

ThoughtDAG prerequisites

• A Windows, macOS or Linux computer.
• Roughly 106–133MB for the listed v0.3.18 installer builds, depending on platform, plus space for the installed app, canvases and any local models.
• Either a supported API key or a locally running OpenAI-compatible model endpoint such as Ollama.
• An official v0.3.18 release asset whose hash matches the table below.

You can explore the example canvas without a provider key. That is worth doing first because it separates “the app launches” from “the model connection works”.

Step 1: download the correct v0.3.18 installer

Open the official v0.3.18 release page and choose the asset for your system. Do not download the .blockmap, .yml or source-code archive unless you know you need it; those files support the updater or source review.

These hashes belong specifically to v0.3.18. If the project has released a newer version when you read this, use the file name and digest shown on that newer official release rather than expecting it to match this table.

Step 2: verify the download

Windows

Open PowerShell and calculate the file hash:

Get-FileHash "$env:USERPROFILE\Downloads\ThoughtDAG.Setup.0.3.18.exe" -Algorithm SHA256

The hash should be:

b2210f184e4326a542bc4bfbdbccf70f7941067f86e75d418ddb4473c402927c

Apple Silicon Mac

shasum -a 256 ~/Downloads/ThoughtDAG-0.3.18-arm64.dmg

Intel Mac

shasum -a 256 ~/Downloads/ThoughtDAG-0.3.18.dmg

Linux

sha256sum ~/Downloads/ThoughtDAG-0.3.18.AppImage

If even one character differs, delete the file and return to the official release page. A matching hash confirms that your file matches GitHub’s published release asset; it is not a substitute for deciding whether you trust the project itself.

Step 3: install ThoughtDAG

Install on Windows

After those checks, open ThoughtDAG.Setup.0.3.18.exe and follow the installer. If SmartScreen appears and you consciously accept the unsigned build, Windows places the continue option under More info. The important distinction is that you verified this exact file first; “Run anyway” is not general advice for unknown downloads.

Install on macOS

Open the appropriate .dmg and move ThoughtDAG to Applications. The project says the macOS builds are signed and notarized by Apple. Apple Silicon machines use the arm64 file; older Intel Macs use the plain .dmg.

Run on Linux

Make the AppImage executable, then start it:

chmod +x ~/Downloads/ThoughtDAG-0.3.18.AppImage
~/Downloads/ThoughtDAG-0.3.18.AppImage

You can move the AppImage to a dedicated applications folder later. Keep the original version in its file name so updates do not become ambiguous.

Step 4: open the example canvas before adding a model

Open ThoughtDAG’s seeded example and learn one rule: only material connected by incoming wires becomes context for the next node. Before spending API credit:

1. Select a response and inspect its incoming context.
2. Open the context row to see messages and estimated tokens.
3. Archive a branch, undo the change and search for a phrase.

If this works, later connection errors are likely in the provider or local-model layer.

Step 5: choose local Ollama or a hosted API

Consumer ChatGPT Plus, Claude Pro and Gemini Advanced subscriptions are not the same as API access. ThoughtDAG’s documentation mentions a community OAuth bridge for ChatGPT plans, but this guide deliberately does not recommend or configure it. Third-party subscription access can breach provider policy or place the account at risk. Use a normal API key, an officially supported plan endpoint, or Ollama.

Step 6: connect ThoughtDAG to Ollama

First install Ollama through its official instructions and pull a model that fits your hardware. Confirm the local service and model list in a terminal:

ollama list
curl http://127.0.0.1:11434/v1/models

Ollama’s official documentation uses http://localhost:11434/v1/ for its OpenAI-compatible API and binds to 127.0.0.1 by default.

In ThoughtDAG, open the model picker and choose the built-in Ollama preset. It uses http://localhost:11434/v1 and does not require an API key. Fetch the model list, then select a model returned by your Ollama server. If you deliberately use the generic custom OpenAI-compatible form instead, enter http://127.0.0.1:11434/v1; some generic clients require a placeholder key such as ollama, which Ollama ignores.

Ask a harmless first question such as “Reply with the model name and one sentence explaining what context you received.” If it works, create a second branch and verify that only the branch you wire in appears in the context preview.

A local model does not automatically make every ThoughtDAG feature offline. Turn off web search and avoid remote materials when the task must stay on the machine. If you want a more reusable local OpenAI-style endpoint on Windows, compare this route with my guide to install LocalAI on Windows with Docker.

Step 7: connect a hosted model provider

Open the model picker, choose a documented provider and paste an API key issued by it. Start with a low-spend model and set a provider budget before loading a long PDF. ThoughtDAG estimates tokens, but the provider controls final price and data handling.

Keys entered through the model picker are stored in browser localStorage and also held by the local proxy in memory. Browser localStorage is persistent device storage, so protect the computer profile and remove saved providers when they are no longer needed. Submitted conversation data still reaches the provider. Check its retention and training settings before using confidential material.

Step 8: build a first useful context graph

Use a decision small enough that you can see the graph working:

1. Create a root question: “Should this small internal API run in Docker or directly as a system service?”
2. Branch once for deployment simplicity and once for operational control.
3. Ask each branch for its assumptions, risks and evidence.
4. Wire the two useful answers into a new synthesis question.
5. Inspect the context preview before sending it.
6. Archive a weak answer instead of deleting the entire history.

The final node should receive the two branches you selected, not every message on the canvas. That visible boundary is the product’s main value.

Step 9: add a PDF without losing track of the source

Drop a PDF onto the landing page or canvas. ThoughtDAG keeps the file local and sends extracted text to the selected model only when you ask about it. Select a passage in the reader, ask a question, and check that the resulting node retains its page reference.

Do not translate “the PDF stays local” into “nothing leaves the computer”. With a hosted model, the extracted passage and relevant context still travel to that provider. With Ollama and online tools disabled, the model request can stay local.

Step 10: turn on backup and hand a graph to a coding agent

Choose a protected folder for automatic backup. ThoughtDAG writes a live <canvas-name>.thoughtdag.json file; Markdown export is simpler when another tool does not need the full graph. Version 0.3.18 documents this existing workflow rather than adding new behaviour.

Start with a read-only coding-agent handoff:

Read architecture-options.thoughtdag.json as decision context.
Summarize the selected conclusion, rejected alternatives and open questions.
Do not edit source files.

For an agent designed to work on a repository, see how to install Prime Agent. ThoughtDAG preserves reasoning context; Prime Agent performs coding work.

If the handoff needs a visual map rather than code changes, use the Diagram Design setup guide for Codex and Claude Code to turn verified nodes and edges into an architecture diagram.

Common ThoughtDAG problems

Windows says “unknown publisher”

The current Windows build is unsigned. Confirm it came from the official release, is named ThoughtDAG.Setup.0.3.18.exe, and matches the hash above. If any check fails, do not run it. Keep SmartScreen enabled.

The hash does not match

Delete it and download it again from the official release. Make sure you did not hash an older copy. A newer version needs the digest from its own release page.

Ollama appears in a terminal but not in ThoughtDAG

Run:

curl http://127.0.0.1:11434/v1/models

If it fails, start Ollama and run ollama list. If it succeeds, check the ThoughtDAG URL for a missing or doubled /v1. Keep Ollama on loopback unless you intentionally secure remote access.

The model ignores an earlier answer

Inspect the incoming edges. A nearby branch is not necessarily in context. Use a full edge for the chain or a reference edge for one result, then confirm the preview before sending.

A hosted provider returns an authentication or model error

Use an API key rather than a consumer subscription login, then select a model enabled for that key. Never share screenshots containing keys.

The PDF opens, but visual page rendering is incomplete

The setup documentation says rendering can fall back to extracted text when an optional dependency is unavailable. Follow the current official setup notes; do not download an unrelated binary.

ThoughtDAG FAQ

Is ThoughtDAG free?

Yes, the app is MIT-licensed. Ollama avoids per-token API charges; hosted providers may charge for model or search usage.

Is ThoughtDAG fully local?

Canvases and dropped files stay in browser or device storage, but processing is local only when you use an endpoint such as Ollama and disable networked tools. Hosted models receive selected graph context and extracted document text. Web and scholarly search contact external services.

Can I use ThoughtDAG with ChatGPT Plus?

Not as a normal API key. ChatGPT subscriptions do not include API access. Avoid third-party subscription authentication; use an official API key, another supported provider or Ollama.

What changed in ThoughtDAG v0.3.18?

No application behaviour changed. The release documents an existing workflow in which automatic backup maintains a live JSON canvas file for a coding agent. It also includes earlier security fixes.

How is ThoughtDAG different from a mind map?

Its wires also control the context supplied to the model; the graph changes the next request rather than merely illustrating it.

Can ThoughtDAG use MCP tools?

Yes, but begin with a plain model connection. Add tools only after reviewing each server’s data access and permissions. See the llama.cpp MCP setup guide for a local example.

Should I use ThoughtDAG or Unsloth Desktop?

Choose ThoughtDAG to structure conversations. Choose Unsloth Desktop to download, run or fine-tune local models through a GUI. They solve different layers.

Should you install ThoughtDAG?

Install ThoughtDAG if you regularly branch into alternatives, revisit technical decisions or need to show exactly why a model received one piece of context and not another. The graph is useful when the reasoning record matters as much as the final answer.

Skip it for short linear chats. The canvas will not improve a one-question task by itself.

For a first serious session, use v0.3.18, verify the installer, choose Ollama for sensitive local work, create two small branches and inspect the context preview before sending the synthesis request. That is enough to learn whether operational context wiring is useful to you without turning the first hour into a provider-and-plugin marathon.

Primary sources

• ThoughtDAG official GitHub repository and README
• ThoughtDAG v0.3.18 release notes and installer assets
• ThoughtDAG v0.3.15 loopback security follow-up
• ThoughtDAG v0.3.14 security fix
• ThoughtDAG setup, providers, cost and privacy documentation
• ThoughtDAG official feature reference
• ThoughtDAG MIT licence
• Ollama official OpenAI compatibility documentation