An integrated research environment — with a gated AI scientist built in.
Literature · design · experiments · notes · Git · LaTeX, on one local desk.
PrismNext is a local-first integrated research environment — not a LaTeX editor with a chat sidebar, and not a generic coding agent pointed at a .tex file.
Everything a research project is made of exists as a first-class object in one workspace: the literature library (Zotero-synced, MinerU-parsed, citation-health-checked), the research brief & plan, experiment workspaces with runs and provenance, reading & writing notes, Git history, and the LaTeX manuscript itself. One enhanced opencode agent runs through all of it — reading papers with you, designing and coding experiments, drafting prose, answering with figures and interactive visuals.
Free and open source, bring your own key. Works with any model provider. There is no Prism cloud, no telemetry, no data collection — projects live on your disk under .prismnext/.
Co-pilot, not autopilot. The agent can advance the work — plan, retrieve, run, draft — but every consequential step passes Plan consent, permission modes, and Proposed Changes review. You keep the veto. That gate is exactly what makes AI assistance admissible in serious research.
| Generic coding agents | Lit-chat / “auto AI scientist” | PrismNext | |
|---|---|---|---|
| Scope | Files + chat + tools | Retrieval Q&A or overnight pipelines | The whole loop: read → design → run → write → review |
| Research objects | None — just text buffers | Chat attachments | Library, brief, plan, runs, notes, manuscript — on disk, queryable |
| Control | Generic approvals | Weak or none | Plan consent, permission modes, Proposed Changes, citation health |
| Data | Cloud / IDE-centric | SaaS-first | Local-first, BYOK, zero telemetry |
| Writing | “Edit the .tex” |
“Here is a generated draft” | First-class LaTeX: Tectonic / TeXLive, live PDF preview, templates |
The loop is the point: ideation, literature, experiments, and writing stay in a single agent-visible workspace — so context never leaks across four disconnected tools.
Screenshots follow your GitHub light / dark mode — the app has five theme packs of its own (see Themes).
Project SQLite library with Zotero sync and MinerU PDF parsing, metadata enrichment (Crossref / arXiv / OpenAlex), BibTeX import / export, and manuscript citation health across .tex ↔ .bib ↔ library. The PDF reader keeps margin notes beside the paper; intensive reading explains a formula at the level you ask for.
Capture the problem and the path in Research Brief and Plan (⌥P), approve the checklist before the build. Experiment workspaces sit beside the chat: plan runs, capture logs and artifacts, and trace which command produced which plot — provenance you can cite in Methods.
A real TeX workspace: Tectonic (bundled) or your TeXLive, % !TEX root / % !TEX program, live PDF preview, Proposed Changes merge view for agent edits, and paper / thesis / beamer templates.
One enhanced opencode agent, many surfaces: multi-tab chat, orchestrator + expert personas, skills / slash commands / project rules, permission modes, and any model provider you bring a key for. It answers in text, figures, and interactive visuals — and it acts on the same library, experiments, and manuscript you see.
Reading and writing notes live next to the work, and every meaningful change can be a commit: built-in Git with worktrees, plus a terminal with AI bash and an in-app browser. The UI speaks English, 简体中文, and 繁體中文（香港）, and the app updates itself.
The agent ships held to a codified set of research standards — 29 bundled skills, each with its reference tables, templates, and runnable scripts. Discipline that must hold every turn lives in an always-on prompt layer; skills load on demand, and every skill stands alone even with all its siblings disabled.
| Family | On board |
|---|---|
| Ideate | idea-lab: divergent brainstorming before judgment · literature cross-pollination · a dedicated ideas/ folder so bold sparks never evaporate |
| Design & run | Hypothesis design · experiment design matrices (cost confirmed before compute) · empirical-ML protocol (seeds, fair baselines, aggregation) · statistical rigor with power analysis · management & decision science (DiD/IV/RDD, behavioral experiments) · run receipts → Methods prose |
| Math | SymPy-verified derivations → LaTeX · numeric probes & convergence checks · differential-geometry and ring/lattice verification — novel claims only |
| Figures | matplotlib templates · Observable Plot data-viz (density, hexbin, facets, geo) → SVG · TikZ/pgfplots vector graphics · reproducible figure pipeline · reopenable panel figures |
| Writing | Outline gate before prose · section skills for Introduction / Preliminaries / Methods / Results / Conclusion / Related Work — patterns, not molds |
| Review | Intensive reading notes · PRISMA 2020 · reverse-angle critical review (backed by an independent peer-reviewer expert) · pre-submission preflight · rebuttal letters |
| Meta | Distill your own skills from workflows that just worked |
Growing with every release — the format is open: connect community skill sources or author your own. Settings → Skills marks each skill's origin (built-in / registry / custom) and lets you disable or delete them individually.
| Mode | How it drives |
|---|---|
| Human-led | You write, run, and compile; ask the agent when stuck |
| Co-drive | The agent advances design, runs, and prose; you approve the major beats |
| AI-led | The agent works the full loop toward the goal; you watch the trail and intervene anywhere |
Autonomy is not the absence of control: every mode keeps the same gates — Plan consent before builds, permission modes for tools, Proposed Changes for edits, and an auditable trail of what ran and why.
- Your machine — manuscript, library, experiments, and agent configuration live in
.prismnext/on your disk - Your keys — model calls go to providers you choose, with your API keys; there is no Prism cloud
- No collection — no telemetry, no analytics, no silent sync; optional network features (metadata enrichment, search MCP, …) are explicit
- Free — Apache-2.0 open source, all platforms
Built for unpublished data, sensitive drafts, and authors who do not want an entire project living only in a SaaS.
Five curated theme packs (Academic · Midnight · Forest · Warm Paper · Graphite), each in light and dark, plus fourteen hand-drawn chat-home backdrops (ink sketch, night rain, starfield, blueprint, …). The screenshots above already follow your light / dark mode — for the full interactive theme tour, visit the download site.
Download macOS, Windows, or Linux (AppImage) from GitHub Releases or the download page.
On macOS, if Gatekeeper reports the app as “damaged,” clear quarantine once:
xattr -cr /path/to/PrismNext.app
Choose a folder as the project root. PrismNext maintains .prismnext/ there (library, brief, experiments, compile cache, …).
Settings (⌘, / Ctrl+,) → pick a provider, paste your API key. Nothing uploads your thesis anywhere.
- Add papers to the library (or search → stage), sync Zotero if you use it
- Capture problem and path in Brief / Plan
- Read intensively, take notes, let the agent synthesize
- Design and run experiments; keep the provenance
- Write in TeX with live preview; review Proposed Changes — keep only what you intend
- Bounded, auditable topic discovery into the local library
- Clearer evidence / stance snapshots from literature synthesis
- Sharper Human / Co-drive / AI-led mode boundaries — autonomy that stays auditable
Issues and PRs that improve the research loop, the consent gates, or the local-first engineering are welcome — see CONTRIBUTING.md, SUPPORT.md, and CODE_OF_CONDUCT.md. Security reports: SECURITY.md.
Media lives in assets/; regenerate the cover with ./scripts/readme-media/generate-readme-cover.sh (see assets/README.md).
Licensed under the Apache License 2.0. Copyright © 2026 yibocat — see NOTICE.
PrismNext — the whole research loop, on your desk, under your gates.