Paper Baker

PDF for you, .tex for your agent.

Paper Baker drops the LaTeX source of every paper you read into your codebase, so your AI coding agent can read and grep the literature with the tools it already has — grounded on exactly the same papers you are.

🌐 paper-baker.web.app · 📦 GitHub Releases

---

Why

You read papers as PDFs. AI agents are native to text. When you ask an agent to implement a method or write a related-work section, it has never actually seen the paper behind it — and a PDF is the wrong format to hand it: equations, tables, and notation only survive in the source.

Paper Baker bridges that gap:

• The CLI (pb) finds a paper on arXiv, downloads its LaTeX source into your project, and keeps a manifest, a refs.bib, and a paperbaker/README.md index up to date — plus a short brief in your root AGENTS.md/CLAUDE.md so agents discover the papers automatically.
• The web app lets you search, organize papers into projects, and read them — the human-facing side of the same library.
• Sync keeps the two in lockstep, so you and your agent share one set of references.

Once the .tex is in your repo, any coding agent (Claude Code, Cursor, …) reads and searches it with its own native tools — no PDF parsing, no lost math.

How it works

pb search "DistilBERT"        →  finds arXiv:1910.01108
pb add 1910.01108             →  downloads the LaTeX into paperbaker/sources/
# now your agent reads it natively:
Grep "\mathcal{L}" paperbaker/sources/distilbert
Read paperbaker/sources/distilbert/main.tex
pb bib > refs.bib             →  a clean bibliography, on tap

Paper sources live in paperbaker/sources/ — a visible directory (so ripgrep/grep and agent tools find it), but sealed inside a nested git repo so the bulky, re-downloadable TeX never lands in your project's history or diffs.

Install the CLI

The pb binary is self-contained, for macOS & Linux:

curl -LsSf https://paper-baker.web.app/install.sh | sh

Windows (PowerShell):

powershell -ExecutionPolicy ByPass -c "irm https://paper-baker.web.app/install.ps1 | iex"

This installs pb into ~/.local/bin and puts it on your PATH.

Quick start

pb login                 # optional — sign in to sync with your account
pb project create        # bind the current directory to a project
pb add 1706.03762        # add a paper by arXiv ID or URL
pb sync                  # download missing sources, regenerate refs.bib

The CLI works fully offline — login is only needed to sync with the web app.

CLI reference

Run pb --help or pb <command> --help for details.

Command	Description
pb search <query>	Search arXiv for papers
pb add <id|url>	Add a paper to the project (downloads tex source)
pb remove <id>	Remove a paper from the project
pb list	List papers in the current project
pb show <id>	Show detailed metadata for a paper
pb read <id>	Print a paper's tex source to stdout
pb context	Concatenate all papers' tex bodies + figure list
pb bib	Regenerate refs.bib and print to stdout
pb sync	Reconcile with the server, re-download sources, regenerate refs.bib
pb project create|list|bind|rename|delete|unbind	Manage projects
pb login / pb logout / pb whoami	Account / auth
pb update / pb uninstall	Manage the pb binary

For coding agents

After a pb add/pb sync, the project contains:

• paperbaker/sources/<paper-id>/ — extracted .tex files, searchable with grep/ripgrep
• paperbaker/refs.bib — a generated BibTeX bibliography
• paperbaker/README.md — a generated guide telling the agent where the sources live and how to search them

pb project create/bind also drops a short, clearly-marked brief into your root AGENTS.md (or CLAUDE.md if that's what the repo uses) pointing at the above — so an agent finds the papers from the repo root without being told. It's written once; opt out with --no-brief.

So you can point your agent at the literature directly:

rg "scaled dot-product" paperbaker/sources/

Repository layout

This is a pnpm monorepo.

Path	Package	Role
apps/cli	@paper-baker/cli	The pb command-line tool
apps/web	@paper-baker/web	React + Vite web app (Firebase Hosting)
functions	@paper-baker/functions	Firebase Cloud Functions (API)
packages/core	@paper-baker/core	Shared logic — arXiv parsing/query, BibTeX, project IDs
packages/providers	@paper-baker/providers	Paper providers (arXiv)
packages/api-client	@paper-baker/api-client	Typed client for the backend API

Development

Prerequisites: Node ≥ 20, pnpm ≥ 9. (Firebase emulators need a JDK.)

pnpm install            # install all workspaces
pnpm dev:web            # run the web app (Vite dev server)
pnpm dev:functions      # run the functions in watch mode
pnpm build              # build every workspace
pnpm lint               # eslint
pnpm typecheck          # tsc across workspaces

Testing

pnpm test               # unit tests (vitest)
pnpm test:integration   # integration tests
pnpm test:rules         # Firestore security-rules tests (emulator)
pnpm test:functions     # Cloud Functions tests (emulator)
pnpm test:e2e           # end-to-end tests
pnpm test:smoke         # full-stack smoke test against the emulators

Run the Firebase emulators locally:

pnpm emulators          # auth + firestore
pnpm emulators:all      # auth + firestore + functions + hosting

Deployment

The web app, installer scripts, and Cloud Functions deploy to Firebase:

pnpm deploy:prod        # pnpm build && firebase deploy --project paper-baker

The pb binary is published as platform builds on GitHub Releases; the install.sh / install.ps1 scripts (served from the web app) fetch the latest.

Tech stack

• CLI: TypeScript, commander, bundled with tsup into a self-contained binary
• Web: React, Vite, Tailwind CSS, HeroUI
• Backend: Firebase (Auth, Firestore, Cloud Functions, Hosting)
• Papers: arXiv source/metadata, throttled to the provider's limits

License

Open source under the MIT License.