A beginner-friendly Codex workspace for mathematical research.
Math Research Workbench helps mathematicians organize ideas, paper notes, proof work, and projects in ordinary Markdown files. It is designed for people who have used ChatGPT but have never used Git, a terminal, or a coding agent. No programming experience is required.
This public repository contains only the reusable framework. It does not contain anyone's research notes, personal paths, credentials, or experimental machine-learning setup.
The recommended route is a release ZIP. GitHub is optional.
-
On macOS or Windows, install the ChatGPT desktop app and sign in. Linux users currently need the Codex CLI and basic terminal familiarity.
-
Download the ZIP from the latest release.
-
Extract it into a local folder that is not managed by iCloud Drive, OneDrive, Dropbox, or Google Drive. A folder named
MathResearchdirectly inside your home folder is usually a good choice. -
In the desktop app, select Codex and open the extracted folder.
-
Send this message:
Start setup
Opening a folder gives Codex access to the workspace, but it does not send a
message or begin setup by itself. You must send Start setup (or invoke
$first-run).
Codex then asks one question at a time about:
- a Python 3.9-or-newer helper runtime for safe saved setup and resume (Codex first uses an existing or bundled runtime; no programming is required);
- your preferred language;
- optional private GitHub backup for text files;
- optional storage for PDFs and other large files;
- optional Obsidian installation and plugins;
- Overleaf or optional local TeX compilation;
- optional Claude Code installation and Anthropic sign-in for Claude reviews;
- optional ChatGPT sign-in inside the in-app Browser for the bundled
$pro-context-bundleconsultation skill, when that Browser is available.
Python 3.9 or newer is used only for safe local setup-state checks; it is not a programming requirement. Every service and editor integration is optional. Codex shows the exact effect and asks for approval before installing software, changing an account, creating a remote repository, or writing outside this folder.
For screen-by-screen instructions and a glossary, read GETTING_STARTED.md.
| Folder | Purpose |
|---|---|
inbox/ |
Unsorted material you want Codex to help classify |
ideas/ |
Questions, conjectures, and evolving research directions |
papers/ |
Bibliographic and reading notes; not copyrighted PDFs |
notes/ |
Reusable definitions, lemmas, examples, and explanations |
projects/ |
Active research, proof work, session notes, and drafts |
files/ |
Local large-file fallback; excluded from GitHub by default |
meta/ |
Conventions, schemas, safety rules, and templates |
You can keep working exactly as you would in an ordinary conversation: ask questions, add information, test approaches, and solve the problem with Codex. For each active project, Codex maintains the project README.md as a compact Research State Spine so that the important mathematical state remains easy to recover.
The spine gives stable, paper- or decision-relevant objects explicit identifiers: Def-NNN for definitions, Lem-NNN for lemmas, Prop-NNN for propositions, Thm-NNN for theorems, Cor-NNN for corollaries, and Gap-NNN for unresolved gaps. These identifiers organize the work; they do not certify that a claim is true. A consequence receives a Cor-NNN identifier only when it has become a stable downstream consequence worth carrying into the paper or later decisions.
Exploratory calculations, temporary observations, failed approaches, and tentative consequences stay in sessions/, logs/, or drafts/. Codex promotes only the stable objects that matter to the paper or a research decision, and you do not need to edit the spine's tables by hand.
For example, you can simply ask:
Keep researching this problem with me as usual. Update the project Research State Spine with any stable definition, lemma, proposition, theorem, corollary, or unresolved gap that matters to the paper, and leave scratch work in the logs or drafts.
The files remain readable without Codex or Obsidian. Obsidian is an optional visual editor for the same Markdown files.
Codex quietly starts a bounded research-run log when a project reaches a second serious attempt, resumes in a later session, needs multiple obligations or tools, or develops a claim that a paper depends on. You do not need a new interface or command. The log fixes the exact problem statement, separates obligations, preserves prove and disprove attempts, records the limited scope of tools or sources, and leaves a clear resume point. If it is unclear whether the threshold has been reached, Codex keeps an ordinary log for now.
Codex does not create this extra record for a simple question, routine
explanation, lightweight exposition, or one-step calculation. It pauses only
when the mathematical statement is ambiguous, an external action needs
approval, independent attempts conflict, or a claim is ready for your explicit
judgment. AI agreement and tool results are evidence with recorded limits, not
automatic proof. Only the researcher can mark a claim closed-by-researcher
or confirm that a checked statement matches the intended mathematics. Each
decision records its date, scope, and a reviewer identifier or role that the
user agrees to include; no email address or other unnecessary personal details
are needed, and an automated check cannot verify someone's identity.
Research-run logs contain unpublished research. Before syncing or pushing, Codex must show what would leave the workspace, obtain approval, and verify that the research destination is exactly private. Never put a run log in the public Math Research Workbench distribution repository.
You can write naturally. For example:
Put this research question into the inbox and help me formulate it precisely.
Read this arXiv link, verify the bibliographic details, and create a paper note.
Audit this proof. Separate established steps, gaps, and possible repairs.
Do not call it verified based only on your own review.
Create a new project for this problem and explain every folder you create.
Keep researching this project with the bounded research-run loop. Preserve
failed approaches and ask me only when a mathematical decision needs me.
See the daily workflow for more examples.
- Unpublished research repositories are private by default.
- Research-run logs must never be committed or pushed to the public Math Research Workbench distribution repository. Verify a separate destination as exactly private before syncing unpublished research.
- GitHub is not an automatic backup. Only changes that have been committed and pushed appear there; ask Codex to explain and confirm each synchronization.
- The workbench repository should not live inside a cloud-sync folder. Use an external drive service for PDFs and other binaries instead.
- Codex never needs your password, access token, private key, or browser cookie pasted into chat.
- Claude Code and Claude are separate Anthropic services. Installation and sign-in each require approval and this bundled review uses only a direct personal Claude Pro or Max subscription login. Team, Enterprise, Console/API keys, cloud providers, proxies, and custom gateways need a different reviewed workflow. Because Claude safe mode does not disable managed policy, the workflow stops when policy is detected or cannot be ruled out and asks the user to confirm the CLI's policy-source screen before each review. Follow the official installation and authentication guidance.
- Signing in is not permission to send research. Before every Claude review or
Browser-based upload or message, Codex must name the provider, show the exact
outbound files, diff, or text, and ask again for that one use. The bundled
$pro-context-bundlewaits for Pro to finish without an overall time limit and never selectsAnswer nowunless the user explicitly asks during that consultation. - Enter passwords, passkeys, MFA responses, and OAuth codes yourself. Codex pauses at credential screens and does not inspect or capture them.
- The project-provided LaTeX delimiter compatibility plugin is bundled inertly and requires separate consent, manual enabling, and a rendering test. Third-party community plugins remain optional; install them one at a time only after reading the plugin guide.
- Obsidian can render mathematical notation without a local TeX installation. Overleaf is the simplest default for compiling manuscripts.
- A language-model review is not a proof. References and final mathematical claims require independent verification.
- A local folder does not make Codex an offline tool. Material needed for a request may be sent to OpenAI for processing. Check your institution's rules before using confidential collaboration, referee, student, patient, or restricted material, and review your account's Data Controls.
On macOS or Windows, use the ChatGPT desktop app for the beginner first-run flow. It can work with a local folder when you grant access. There is currently no Linux desktop app; Linux users need the Codex CLI, which assumes some terminal familiarity; follow the current official CLI guide. A hosted Codex/cloud checkout cannot install Obsidian or TeX on your computer, inspect a local cloud-drive folder, or preserve this machine's ignored setup file. In that situation the workbench remains usable as Markdown, but Codex will point you back to the desktop setup guide instead of pretending local setup is complete. See OpenAI's current desktop and Codex overview.
The optional in-app Browser is available only on supported macOS or Windows
desktop configurations, and its availability can depend on product capability,
plan, or workspace policy. It is unavailable on Linux. Its profile is separate
from your ordinary browser and other app sessions, so ChatGPT may ask you to
sign in again. The bundled $pro-context-bundle uses it; choose later if you
do not plan to use that skill. See the official
Browser guide.
On macOS or Windows, the Codex CLI is supported for experienced users but is not required for the beginner desktop route.
- Detailed setup
- Obsidian guide
- Troubleshooting
- Updating safely
- Security policy
- Contributing
- License and content notice
Math Research Workbench is an independent project and is not affiliated with OpenAI, Anthropic, ChatGPT, Codex, Claude, Obsidian, GitHub, Overleaf, or any cloud-storage provider.