An AI assistant for mathematics
Work on mathematics with Prismo, an AI assistant that reads your definitions, cites your sources, proposes proof edits and runs Python checks.
It reads before it answers
The assistant in Manifield is called Prismo. It works from your project rather than from memory: it searches your library, reads the items it needs, and cites them by id. When you ask about "the closed graph theorem", it opens your item thm-closed-graph, with your definitions and your notation, before it says anything.
It always has your notation file and your research memory (memory.md, the project's notes on goals, what was tried and what failed) in view, and it keeps the memory up to date as you work.
Every mention is a link
When Prismo mentions an item, the id becomes a link. Click it and the item opens beside the conversation, rendered as in the library, so you can check a statement without losing your place. Items from an attached book or paper are linked the same way, as source:item-id.
You decide what gets written
The composer has four modes:
| Mode | What Prismo may do |
|---|---|
| Ask | Read and discuss. No changes |
| Plan | Draft a research target for research mode |
| Propose (the default) | Everything, but every change to the library comes to you as a proposal you accept or reject |
| Auto | Write to the library directly. Manifield asks you to confirm before switching it on |
Statuses have a hard limit in every mode: an agent can raise a result as far as Review found no errors. Human approved, Cited literature and Formally verified are set only by you.
It can compute
Ask Prismo to check a conjecture and it writes a Python script, using networkx, sympy or numpy, and runs it in a sandbox on your Mac:
- No network, confined writes and a hard time limit, enforced by a macOS sandbox profile.
- Every run is kept. A folder under
artifacts/holds the script, its output, its errors and a record of the run: the evidence behind any "I checked every graph up to seven vertices" claim. The chat's Artifacts button opens it. - Results attach to items as evidence. A clean run can move a conjecture to Tested; a counterexample sends it back to Conjecture. Like every write, this is a proposal by default.
It can also look integer sequences up in the OEIS.
It can draw
Prismo can render a picture with Python, and keep it: a saved figure becomes an item in your library, with its caption, what it illustrates, and the script that drew it. Agents cannot see images, so a figure's caption is the only description any of them gets. Manifield insists on one.
It can use your sources
If you have attached books or papers to the project (sources), Prismo searches them and cites their own statements. It reads a full text only with your permission: it asks in the chat (Allow or Deny), or you switch on standing access in the composer. The same holds for documents kept in the project.
What it does not do
- It does not browse the web. The chat can query the OEIS. Research mode can search arXiv for titles and abstracts while it plans.
- It does not prove things "correct". The best outcome of an AI review is Review found no errors. Manifield records what evidence exists, never more.
- It does not run on your Mac. Models run at the provider you choose; the context sent to them (the items Prismo reads, your notation, your question) goes to that provider.
Other agents in the app
- The Organizer turns an imported
.texfile or PDF into proposed items. - The Explorer suggests new conjectures and connections from the library and graph toolbars. It is the least tested agent in the app.
- The Prover and the Reviewer draft proofs and hunt for flaws in them, inside research mode.
Each can run on its own model: Settings → Roles.
Try it on a proof draft
The included Baire notes contain an unfinished step in the closed graph theorem. The worked example follows the definitions, dependencies and norm estimate you can check. For a broader workflow, see studying advanced mathematics with AI.