Marks of your own
Comments C and bookmarks B sit in the margin next to thread markers. They're yours: the agent never receives them.
Early access · free on Windows · macOS and Linux soon
Go down the rabbit hole. Keep your way back.
Qrabbit is a reader for difficult material. Highlight a passage, explore it with AI, branch off inside any answer and go as deep as you need. The paper stays where you left it, and everything you explored stays attached to the passage it came from.
Runs on your Claude or ChatGPT subscription, through Claude Code or Codex. No API key.
No more copying text. Just highlight it.
Select a passage and choose Explore. The thread opens beside the paper with the passage and the text around it already in hand, and the highlight stays in the paper as your way back to it.
A typical chat
So how do the four value functions relate?
The optimal ones are the best the on-policy ones can be, V*(s) = maxπVπ(s), and the same holds for Q. Each can be written in terms of the other…
CtrlC in the paper → AltTab → CtrlV
In deep RL, we deal with parameterized policies: policies whose outputs are computable functions that depend on a set of parametersWhat does this mean?
Which text is this from? Without the paragraphs around it, I can only guess what the parameters are.
It's from the Spinning Up page I pasted earlier. Section 3, I think?
Qrabbit
AI loves lists. Keep them tidy.
When an answer has several parts, AI reaches for a list, and following up on each point turns one chat into a tangle of lists inside lists. In Qrabbit you highlight a point and it opens its own branch. The outline keeps every branch in order, and the answer stays where it was.
A typical chat
- Actor-critic methods combine the two…
- Model-based methods plan with a learned model…
Go deeper on 3.
It comes down to three ideas:
- The actor picks actions from the policy…
- The critic scores them against a baseline…
- The advantage says how much better than usual…
Why does 2 cut the variance?
Do you mean the critic here, or value-based methods, point 2 in the first list?
The critic. And what was 4 in the first list again?
↑ The first list is three answers up
Qrabbit
The rabbit hole
A rabbit hole isn't the problem. Losing the way back is.
Curiosity should be cheap. Follow it as deep as it goes, keep exactly what you learned, and come back to the sentence you were reading.
- What are the main ways to learn a good policy?threadincluded 1 branch
- Actor-critic methodsbranchhandoff from parent
- the advantagebranch
- same expectation, far less spreadbranch
- PPObranch
- the advantagebranch
- Model-based methodsbranch
- Actor-critic methodsbranchhandoff from parent
- 1
Highlight and Explore
Select a passage and press E. A thread opens beside the paper, anchored to it, with the text around it already in hand.
- 2
Explore inside the answer
Select a phrase in the answer and Explore again. The branch opens with a handoff from its parent, so it doesn't start from scratch.
- 3
Keep going
Branch inside the branch, and again. Every tunnel stays on the map: the breadcrumb, the outline and the markers in each answer.
- 4
Head back up
Click the parent or press Alt+↑. There's nothing to close or resolve: the parent picks up what the branch learned with your next question.
Also in the box
Made for reading, not for chatting.
Web pages read like papers
Add a URL and it's fetched once, reduced to its article text and saved as Markdown, so you can highlight it, explore it and read it offline.
Maths and code, intact
Formulas published as images, MathML or MathJax come back as real LaTeX. Code blocks keep their language, and callouts stay callouts.
Answers with the page in hand
The agent answers with the passage, the text around it and page numbers, and reads the rest of the source itself when it needs to.
Out of the way
No sidebar. Collections, sources, threads and marks live in header menus, the divider drags, Focus hides the paper, and everything has a shortcut.
You approve what it touches
Reads inside the collection and writes under notes/ are allowed. Anything else asks you first, right in the thread.
Collections
One reading list per subject. One plain folder per list.
Put sources in the order you'll read them, tick off what you've finished, and pick up where you left off. Reading positions are kept per source and per thread.
Files are the truth.
A collection is an ordinary folder of Markdown, JSON and PDFs. Open it in Obsidian, track it with git, back it up however you like. The app's own state lives in .qrabbit/ and is safe to delete; removed sources wait a week in its trash.
reinforcement-learning/
qrabbit.json
the manifest and your reading order
AGENTS.md
instructions any agent reads here
sources/
PDFs, and web pages saved as Markdown
key-concepts-in-rl/
source.md
the article, with its link
text.md
extracted text with page markers
annotations.json
your comments and bookmarks
threads/
one Markdown file per thread
notes/
where the agent may write
Bring your own subscription
Use the AI you already pay for. Or run your own.
Qrabbit doesn't sell AI and never asks for an API key. It drives the coding agent you already have through the Agent Client Protocol, so every turn comes out of your existing plan. Claude Code and Codex work today; more agents and local models are on the way.
Claude Code · connected
Key Concepts in RLIn deep RL, we deal with parameterized policies: policies whose outputs are computable functions that depend on a set of parameters.
Ask a follow-up, or highlight the paper and choose Ask here
Claude MaxmodelOpuseffortHigh
Sign in once through Claude Code's own terminal flow, started from the header menu.
- Qrabbitspeaks ACP
- The agent's own appsigns in its own way
- Your planpays for the turn
Qrabbit never reads an agent's tokens, config files or keychain entries.
Pricing
Free to read. Cloud if you want it.
The app is a free download. Qrabbit Cloud is for keeping your work safe and in step across machines.
Qrabbit
$0free download
The whole reader. Your files, your agent, your machine.
- Highlight, explore and branch as deep as you like
- PDFs, Markdown and saved web pages
- Unlimited collections, as plain folders on your disk
- Runs on the agent subscription you already have
Qrabbit Cloud
$5per month
Your work, saved and synced. Pick up on any computer where you left off.
- Collections, threads and marks synced between your computers
- Everything saved off your machine, so a lost laptop isn't lost work
- Still plain files locally: the cloud keeps a copy, never the only one
- AI still runs on your own subscription; Cloud doesn't resell it
Download
Pick something hard to read. Start there.
Qrabbit is in early access and free. To power the agent you'll need a Claude plan that includes Claude Code, or a ChatGPT plan that includes Codex; you sign in from inside the app.
- Needs a Claude Pro or Max plan for Claude Code, or a ChatGPT Plus or Pro plan for Codex.
- Keeps your collections in folders you choose, on your own disk.
- Updates arrive as new releases while Qrabbit is in early access.
Questions
Before you dig in
Do I need an API key?
No. Qrabbit drives Claude Code or Codex through the Agent Client Protocol, and each signs in with your own Claude or ChatGPT plan in its own flow. Qrabbit never sees a credential.
Which subscriptions work?
Today, any Claude plan that includes Claude Code (or an Anthropic Console account), and any ChatGPT plan that includes Codex. GitHub Copilot, Gemini through Antigravity, other ACP agents and local models are planned.
Where is my work stored?
In a plain folder per collection: the sources, one Markdown file per thread, your marks as JSON. Open it in Obsidian or track it with git whenever you like.
Does the agent see my comments and bookmarks?
No. They're for you, and they never go into a prompt. The agent sees the passage you highlight, the text around it and what you ask.
What does Qrabbit Cloud add?
Sync and backup of your collections for $5 a month, when it launches. The app itself stays free, and your files stay on your disk either way.
What can I read with it?
PDFs, Markdown and text files, and web pages, which are saved as clean Markdown so they read like papers and work offline.