PrismNext
Early Access

Preprint Version …

PrismNext

A collaborative AI scientist on your desk.

Abstract

PrismNext keeps read → design → run → write → review in one local workspace: a project literature library, gated experiment runs, and first-class LaTeX. The agent advances the work; you keep the veto.

Keywordslocal-first · LaTeX · literature · experiments · gated agent

macOS .dmg Windows .exe Linux AppImage

Local-first · Bring your own API key

co-drive, not autopilot →

§1 The research loop

Ideation, literature, experiments, and writing stay in the same local Agent surface — not four disconnected tools.

  1. 01
    Ideation

    Problem and path, captured in Brief and Plan.

  2. 02
    Literature

    A project shelf with citation health — not chat attachments.

  3. 03
    Experiment

    Runs, logs, and provenance for Methods.

  4. 04
    Writing

    TeX workspace, live PDF, Proposed Changes review.

Fig. 1. The research manifold — drag to rotate.
read → design → run → write

§2 Capabilities

Evidence over adjectives — nine surfaces that ship today, each wearing the theme pack you pick in the header.

One prompt box, the whole desk

A new Agent session starts here: pick a model, an expert, and skills, then aim the same composer at a literature hunt, an experiment run, or a manuscript.

Light Dark
Fig. 2. The session composer — model, expert, and skills at hand.

Literature, searched and shelved

The agent hunts across sources and stages what it finds — Pending vs In library is explicit, Zotero syncs in, MinerU parses the PDFs. Papers live in the project, not in chat attachments.

Fig. 3. Search results with their shelf status, one click from the library.

Reading with a co-pilot

The PDF sits on one side, the agent on the other. Ask about a section, a figure, or a claim — answers cite the page you are looking at.

Fig. 4. The paper and its reading companion, side by side.

Intensive reading, line by line

Lasso a formula, a figure, or a paragraph and the agent walks through it step by step — notation unpacked, derivation explained, assumptions named.

Fig. 5. A lassoed formula, explained line by line.

Notes that write back

Derivations, reading cards, and half-ideas live in project notes — and the agent works inside them: expanding, organizing, and cross-linking as you think.

Fig. 6. Notes on the left, the agent expanding them on the right.

Experiments with provenance

The same agent writes the script, runs it gated, and files the receipt — command, exit code, runtime, run id, output, artifacts. Methods-grade traceability by default.

Fig. 7. A finished run with its full receipt.

Git, built into the workspace

Every step — agent-written or yours — lands in git. Side-by-side diffs, commits, and reverts without leaving the desk; the whole research trail stays diffable.

Fig. 8. The day's work as a side-by-side diff, ready to commit.

First-class LaTeX writing

A real TeX workbench: source with an outline sidebar, live PDF preview, bundled Tectonic compile, and Proposed Changes review — serious manuscript editing, not a Markdown sidebar.

Fig. 9. Source, outline, and the freshly compiled PDF.

Every provider, your keys

DeepSeek, Claude, Gemini, GPT, Grok, Kimi, Qwen, MiMo, MiniMax — switch models mid-task, all on your own API keys. No Prism cloud, no middleman.

Fig. 10. The model picker — every major provider under your keys.

Interactive research, with a human gate

The agent proposes, you dispose: plans ask for consent, large moves come with auditable diffs, and permission modes decide how far autonomy goes — co-drive when you want control, full loop when you don't.

Fig. 11. A working session — the agent acts, you review every large move.

Research standards, codified

A small set of research standards the agent is held to — each with the tables, templates, and runnable scripts to enforce it. Open format: plug in community sources or write your own.

Orchestration · 10

  • Project kickoffIdea → structured project: brief, plan, skeleton.
  • Related-work pipelineDiscover → screen → stage → cite, on the literature tools.
  • Intensive reading notesStructured deep-read notes from PDFs.
  • Hypothesis designSharpen a question into testable hypotheses.
  • Experiment design matrixFix the factorial / ablation matrix before running.
  • Experiment-to-MethodsExperiment logs become Methods prose.
  • Figure pipelineData → figure → manuscript, end to end.
  • Critical self-reviewClaims-vs-evidence structured self-critique.
  • Rebuttal letterPoint-by-point reviewer rebuttal drafting.
  • Manuscript preflightPass/fail gate before sharing: compile, citations, desk-reject.

Method & rigor · 6

  • Statistical rigorTest selection, effect sizes, runnable power analysis.
  • PRISMA systematic reviewPRISMA 2020 protocol, screening log, flow counts.
  • Scientific visualizationColorblind-safe matplotlib style + chart selection.
  • Symbolic mathSymPy-verified derivations, straight to LaTeX.
  • TikZ & pgfplotsTikZ/pgfplots templates that compile out of the box.
  • Panel figuresRightArea figure & plot objects, conventions included.

Discipline protocols · 2

  • ML experiment protocolMulti-seed discipline, fair baselines, aggregation script.
  • Management & decision scienceDiD/IV/RDD, behavioral experiments, robustness battery.

Meta · 1

  • Skill creatorAuthor and validate your own skills.
all local — check .prismnext/ figures wear the current pack → runnable, not rhetoric ↓

§3 Axioms

Axiom 1 (Locality). Your projects stay on your machine.

Axiom 2 (Privacy). No telemetry, no analytics — nothing leaves the machine.

Axiom 3 (Keys). Use your own model API keys.

Axiom 4 (Veto). Every automated move is gated and auditable; you keep the veto.