I actively help the mathematical research community adapt to modern AI tools. As part of this effort, I develop and provide MathInery, scaffolding for AI-assisted mathematical research, along with some of my harness tooling.
If MathInery contributes meaningfully to your work, please acknowledge and cite it, preferably using its Zenodo DOI. Every decision not to acknowledge community-maintained infrastructure is a silent vote for dependence on commercial providers.
Downloads
If you are unsure where to start, use the deployed MathInery package. It includes ready-to-use configurations for Claude Code, Codex, OpenCode, and Prime Agent and does not require Harness Bootstrap. The Zenodo package is also provided in deployed form.
My LaTeX notes templates are available separately if you would like a starting point. See also the "Adjustment to personal use" section in MathInery's README.
For more customization or to develop your own harness infrastructure, use the packages below. Deploying MathInery or MathInery Literature from source requires Harness Bootstrap.
- Harness Bubblewrap (2026-11-10).
- Harness Bootstrap (2026-11-10).
- MathInery (2026-11-10).
- MathInery Literature (2026-11-10).
Harness Bubblewrap: Make it safe
Harness Bubblewrap installs a family of Bubblewrap shell commands that sandbox agent harnesses on Linux. They retain network access and read-only access to basic infrastructure, while the mounted project and harness state remain writable.
Consider the level of isolation you want to work with:
- Yolo harness: bypass approvals and built-in sandboxing, running commands with your account's access; highly unsafe but convenient.
- Usual harness: retain built-in permissions, approvals, and sandbox where available; trust the provider's implementation, despite a sometimes questionable track record and incentives.
- Extra sandbox + yolo harness: remove harness checks while retaining the outer sandbox.
- Extra sandbox + usual harness: restrict filesystem access independently of the harness and keep its checks.
- Docker or Podman: isolate a container, sharing the host kernel; rootless mode reduces privileges but can complicate file permissions.
- Virtual machine: isolate a guest kernel and system for stronger separation, with more setup; shared folders remain accessible.
- Air-gapped system: disconnect from networks, requiring local models and offline tools; you probably already know if you need this level of security.
Whichever level you choose, consider also the 3-2-1 backup strategy with encrypted backups in of the unexpected:
- 3 total copies of your data (1 original plus 2 separate backups);
- 2 different types of storage media (internal SSD, external hard drive, cloud, etc.);
- 1 copy stored off-site (e.g., cloud storage or a physical drive kept at a different location).
Harness Bootstrap: Keep it flexible
Harness Bootstrap is a script that translates a shared infrastructure into the native formats used by Claude Code, Codex, OpenCode, and Prime Agent. Prompts, agents, skills, commands, and model choices are maintained in a single source repository, from which the script generates the configuration for each harness.
AI providers tie their subscriptions to the use of their own harnesses, creating vendor lock-in that goes beyond the quality of the models and harnesses themselves. Harness Bootstrap keeps a further layer of your workflow portable across harnesses. This portability has limits: different models and harnesses benefit from tailored instructions and settings, and some features have no direct equivalent. Even so, the added flexibility currently seems worth the effort.
The package also includes a basic AGENTS.md, the harness-optimizer agent, and skills to help develop and refine your own infrastructure. Keeping the source definitions separate from their deployed copies lets you develop agents and skills independently of those currently in use, then regenerate the harness configurations when you are ready to apply the changes. Together, the optimizer and generator support a partly automated improvement loop that helps you adapt your harness infrastructure to your workflow. You can fully automatize this workflow inside of a container or virtual machine to obtain a reuseable self-optimized harness.
MathInery: AI is a tool. You are the operator.
A researcher-operated system for AI-augmented mathematics.
MathInery coordinates specialist agents for mathematical research, writing, and review. It keeps source readings, calculations, proofs, and review findings alongside the documents they concern.
A research workflow may begin with an incomplete idea and proceed through literature search and exploration to exposition and verification.
Document review is central to MathInery. Mathematical publications require correct arguments and clear, concise exposition. Separate specialist agents examine the mathematics and the exposition of documents written by researchers or generated by models.
Intended use
The researcher operating MathInery retains intellectual authority. They supply the mathematical context, select the approaches worth pursuing, and accept, edit, or reject proposed text. Between these decisions, bounded operations proceed independently to the extent the researcher requests.
Traceable records
Models may invent citations, overlook relevant literature, suggest false lemmas, or produce persuasive but defective arguments. MathInery addresses these risks with traceable sources, tests, independent review, and escalation to the researcher.
Each specialist agent receives a task with explicit inputs, permitted files, outputs, and completion checks. Plans, intermediate artifacts, checks, failures, and open questions form a durable research record from which work can resume. The record can also serve as the basis for systematically documenting AI contributions.
MathInery Literature: AI for legwork. You read for insight.
MathInery Literature helps tracking ongoing developments in the literature and supports learning from individual papers, possibly across sessions.
The included paper tutor reflects some initial ideas but does not yet work well. It is provided mainly for completeness.