v0.3.0-beta.2

Beta 2: the desktop now notices new releases. A sticky notice appears in-app within half an hour of a new version, linking to the download page. Also carries every web-side improvement since beta.1 (turn version pairs, material wires solid by default, the four reading-session fixes, first-launch tutorial, splash screen). Still unsigned; see beta.1 notes for the first-launch steps.

v0.3.0-beta.1 — ThoughtDAG gets a window of its own

The desktop app is here, in beta: the same ThoughtDAG, in its own window, with the local server bundled inside. No Node, no terminal, no setup — download, open, think.

What you get

• macOS (.dmg — Apple Silicon & Intel), Windows (.exe), Linux (.AppImage)
• The full v0.2 feature set: reading loop, clipping, condense, one-click OpenRouter sign-in, subscription presets, keyless local web search
• A local-first home: your canvases live on your machine, keys stay in the app, nothing touches a server of ours
• First launch opens the three-minute lesson; a splash screen keeps you company while the engine warms up

Beta honesty

• Builds are not yet code-signed. macOS: right-click the app → Open on first launch. Windows: SmartScreen → "More info" → "Run anyway". Signing arrives with the stable release.
• Auto-update is not wired yet — beta versions are replaced by downloading the next one.
• Found an edge? Issues are very welcome.

The web app remains exactly one URL away, unchanged: https://app.thoughtdag.workers.dev

v0.2.0 — Clip it, condense it, connect in one click

The one rule hasn't moved: a wire is context. This release is about what flows along those wires — how material gets in, how a grown map slims down, and how a model gets connected without ceremony.

Read, clip, capture

• Clip from the reader. Select a passage in any document and clip it to the canvas as a note; drag a rectangle over a page and the area becomes an image node. Every clip carries a provenance chip that jumps back to its exact source.
• Images are first-class material now. The reader shows them full size with zoom and drag-to-pan; a vision pass reads them into editable companion text (any text inside gets transcribed); the same rubber-band capture clips figures out of them — always from the original resolution.
• Highlight the source itself. Passages you mark in a document render highlighted in its text view, alongside the highlights you make in answers.

Condense: compaction you can watch

Long conversations grow corridors of refinement turns. Condense scans for them (a pure graph walk — no model decides what is collapsible), shows each run with its token saving, and builds a condensed copy beside the original. The original tree is never touched.

• Decision and pivot turns are landmarks: they never collapse.
• Your highlights survive verbatim — machine prose compresses, human judgment doesn't.
• The last steps before every leaf are the working frontier and stay out.
• Every distill node carries a chip back to the run it replaced. Wire the same question into both trees and compare the answers; archive the original when the copy has earned it.
• Builds run in the background — close the window, keep working, watch the toolbar pulse.

Connect without ceremony

• Sign in with OpenRouter. One authorization mints a key inside your browser — nothing to paste, free-tier models included, and the key still never touches any server of ours.
• Subscription plans as first-class doors: GLM Coding, Kimi Code, and MiniMax coding plans connect through dedicated presets; the ChatGPT plan connects through a community local bridge, now with an honest note about whose account takes the risk.
• DeepSeek streams straight from the browser on the hosted app — thinking models no longer die of the proxy's CPU allowance mid-thought.
• Web search without a key. Local runs search out of the box on an anonymous aggregator tier; bring a key only if you want a bigger daily quota. The hosted app spends only keys you brought.
• Windows in the open. Model context windows show in the picker; when an endpoint under-reports its own metadata, the model's name wins upward — and when vision reroutes a request, it says so, names who actually answered, and rescues once with companion text if the stand-in fails.

Wires and versions tell the truth

• A version is a (question, answer) pair. Editing a question no longer orphans the answers written against the old wording — the version arrows switch the whole turn, and downstream staleness reacts to whichever pair is active.
• Material wires solid by default. Connecting a file, note, or link now wires its full content in (with the real token price shown), because a summary reference from a material was a nearly-empty gesture that looked like "file connected".
• Every material card confesses its wiring: not wired · summary only · wired in full — the card tells you exactly what the model will read.

Around the canvas

• The corner minimap magnifies on hover, wears type colors (materials, branches, distills, archived), and zooms with the wheel.
• The palette does click and drag again; pasted files beat their own path text; fast thinkers no longer flood the canvas (stream frames batch); map-view double-clicks open documents where they are read.
• First-run courtesies: the landing quick-connects by language, Escape always closes, and nothing renders an empty map.

Getting started

The live demo runs in the browser — no install, no signup. Load the example canvas, or drop a PDF and start from reading.

git clone https://github.com/chenxiachan/thoughtdag && cd thoughtdag
npm install && npm run server & npm run dev

Any OpenAI-compatible endpoint works, local runtimes included. Keys live in your browser and your machine's memory, never on disk, never on our servers.

Known edges

• Condense distills with your selected model; heavy thinking models are slower to distill — a fast model does it better.
• The anonymous search tier meters per IP per day; the exact quota is the provider's to change.
• Directory-based auto-backup remains Chromium-only (Safari and Firefox export manually).
• Desktop packages are in the works — this release is the web foundation they will ship with.

v0.1.0 — the map is open

The first public release of ThoughtDAG: an infinite canvas where LLM conversations grow into an editable thought graph.

The one rule

Wires are the context. What the model sees is exactly what wires into the node. Editing the graph edits the model's memory — and you can test it: delete one edge, ask again, and the same prompt returns a different answer.

What's inside

• Canvas & context: continue/explore/reference edges with per-edge token pricing, quote ⇄ full conversion, archive, merge synthesis, a "will send" preview of exactly what the model will read
• Reading loop: original-page PDF reader; select a passage and ask, the answer lands on the canvas with its page number; guided digests as wireable nodes; scanned pages recognize into Markdown
• Highlights: mark what matters, review every mark in one overview, weave any subset into one cited passage
• Map view: three semantic-zoom tiers (cards → takeaway plaques → glyph seals), every step badged by its cognitive move (✕ ⚖ ↩ ?), plus a creation-time rail and a journey paragraph
• Staleness & replay: upstream edits mark the answers they invalidate; replay in dependency order, token estimate first
• Paradigms: human/machine workflows saved as files; change the input, replay the experiment
• Models: nine provider presets or any OpenAI-compatible endpoint; per-node pins that follow the line; local Ollama runs fully offline; free tiers (Z.ai/Zhipu GLM) cover every feature
• Local-first: everything lives in your browser; automatic folder backup writes real files; share links carry the whole graph in the URL, no server storage; a Gedankengang poster export prints the journey

Getting started

• Hosted demo (the example canvas needs no key): https://app.thoughtdag.workers.dev
• Local: npm install && npm run server & npm run dev — details in docs/setup.md

Known edges

• Automatic folder backup needs a Chromium browser (Safari/Firefox: use one-click manual export)
• PDF text extraction runs on the local proxy only; on the hosted demo the file still attaches but extraction is skipped
• The mobile layout is browseable but the toolbar overflows on narrow screens; a desktop browser is the intended home
• Consumer subscriptions (ChatGPT Plus etc.) do not include API access; see docs/setup.md for the free key paths

If ThoughtDAG plays a role in your research, the repository ships a CITATION.cff (and this release carries a Zenodo DOI).

The graph is acyclic. You are the loop.