The all-around AI-agent research collaboration app — an embedded Pi agent with Teams v2 on a local-first desk.
Brainstorm · literature · experiments · critique · writing — run the loop autonomously or together.
Website · Changelog · Releases
PrismNext is an all-around AI-agent research collaboration app — not a LaTeX editor with a chat sidebar, and not an overnight auto-scientist.
You set the direction; the embedded Pi agent and Teams v2 do the rest at whatever autonomy you choose: run the full research loop on their own — read, think, critique, experiment, write, review — or work shoulder-to-shoulder as your copilot. Either way, every step stays gated, observable, and local-first.
What makes it different:
- Teams v2 — switching the team switches the working mode. A team is not a prompt persona: it staffs the desk with one Lead voice, specialist subagents delegated via Task, skills, slash commands, and MCP tools. Core ships 8 built-in teams plus 30 research skills; Pro adds debate arenas, mock committees, and career coaching.
- One desk for the whole loop. Literature, briefs, plans, experiments, notes, Git, and the manuscript are first-class objects in the app — not scattered across a reader, a terminal, and an editor.
- Autonomy with a leash. Interactive Plans (⌥P), permission modes, Allow/Deny cards, and live Job Monitor: run fully autonomous or stay in the driver's seat.
- Provenance you can cite. Every experiment run receipts its command, exit code, duration, and artifacts into
runs.jsonl— Methods paragraphs trace back to real runs. - Local-first, BYOK, zero telemetry. Your machine, your keys, your data. Remote Hosts over SSH get the same sealed-key treatment (AES-256-GCM, unwrap key never leaves this computer).
A team is not a prompt persona. It staffs the desk: one Lead speaks in chat, specialist subagents arrive via Task as child sessions, and each team carries its own skills, slash commands, and MCP tools. Core bundles 8 teams (research, writing, figures, review, math, …) that reshape how the agent plans, critiques, and writes; Pro adds debate arenas, mock committees, and career coaching. Switching the active team restaffs the session in place — the same composer, a different way of working.
Several paper folders stay open at once — each with its own chats, file tree, library, and modes. Switching focus never kills background agents; a chat waiting for approval surfaces a title-bar chip so you can jump back when ready.
Per-project SQLite libraries with two-way Zotero sync, Crossref / arXiv / OpenAlex / Semantic Scholar search as agent tools, optional MinerU parsing, and continuous citation health auditing (.tex / .typ ↔ .bib ↔ library). Papers found mid-chat land in a side panel — one click shelves them into the library. Office documents, EPUB, and CSV convert locally via AnyDoc — no API key.
Open the paper and its co-pilot side by side: ask about a lemma, a claim, or a plot and get answers with page citations.
Research Brief and Plan capture intent before compute; experiment runs are gated by permission modes, monitored live, and receipted into runs.jsonl. Long-running jobs keep going when you close the tab.
The pane is live — Typst renders as you type, and LaTeX compiles to PDF in the background. Write alone or let the agent draft; every edit arrives as a reviewable Proposed Changes diff. An in-paper composer lets the agent revise the manuscript directly where you are reading it.
Browse every bundled skill — its protocol, templates, and checks — from one place.
Open a lab machine from ~/.ssh/config. The Host runtime installs itself; chat, files, literature, compile, and experiments run on the server while your laptop stays the control desk. Remote sessions keep an offline laptop copy you can browse on cold start.
| Dimension | Generic coding agents | Literature Q&A / auto-scientists | PrismNext |
|---|---|---|---|
| Research loop | Fragmented across IDE, terminals, and chat | Overnight runs with unchecked output | One continuous loop — ideate → read → think → run → write → review — at your chosen autonomy |
| Agent paradigm | A single generic chatbot | A fixed prompt persona | Teams v2: Lead + Task-delegated specialists + skills + MCP, one switch to restaff |
| Literature | Ignores the library | Chat-with-PDF only | Fine-grained survey: per-project library, Zotero sync, lasso-a-formula intensive reading, citation health across .tex/.typ ↔ .bib |
| Reasoning | Free-form chat | Hallucination-prone essays | Symbolic math (SymPy-verified), derivation checks, critique passes against evidence |
| Experiments | Ephemeral unmonitored subshells | Blackbox cloud VMs | Gated runs, live Job Monitor, runs.jsonl receipts citable in Methods |
| Remote work | Dev containers at best | Fully hosted blackbox | Zero-touch deployment over SSH: Host self-installs runtime, laptop stays the control desk, sealed keys |
| Artifacts | Plain text buffers | Chat attachments | Library, Briefs, Plans, Runs, Notes, Manuscript — all real objects on disk |
| Writing | Markdown or naive text editing | Unchecked generated text | First-class LaTeX & Typst, live preview, Proposed Changes diff |
| Governance | Loose permission prompts | None | Plans, permission modes, Allow/Deny cards, human veto |
| Privacy | Cloud-dependent / SaaS | Remote hosted servers | Local-first, BYOK, zero telemetry |
The default Core team ships 30 codified scientific skills — protocol tables, LaTeX/Typst templates, and executable checks:
| Domain | Skills |
|---|---|
| Ideate & design | idea-lab · hypothesis-design · experiment-design-matrix · ml-research-protocol · statistical-rigor · management-science-empirical · experiment-to-methods |
| Write | writing-design · writing-introduction · writing-preliminaries · writing-methods · writing-results · writing-conclusion · writing-related-work |
| Figures | figure-matplotlib · figure-observable-plot · figure-tikz · figure-typst · figure-pipeline · figure-interaction |
| Read & review | intensive-reading-notes · prisma-systematic-review · critical-review · manuscript-preflight · rebuttal-letter |
| Math & meta | symbolic-math · math-numeric · math-manifold · math-lattice · skill-creator |
Beyond Core, the official beta bundles eight optional Pro teams for high-stakes milestones: Idea Arena (structured debate) · The Committee (mock defense) · Rebuttal War Room · Milestone Coach (career timeline) · Claim Police (claim–evidence audit) · Translation Table (cross-discipline) · Topic Brainstorm · Idea Ledger.
Activate: in a Pro-enabled build, go to Settings → About, paste
PRISM-PRO-DEV-TEST, and click Activate — free during Early Access.
- Locality — manuscripts stay in your Git tree; app state lives in
~/.prismnext/, project metadata in<project>/.workbench/. Nothing uploads to a PrismNext cloud. - BYOK — model calls go directly from your machine to your provider. No proxy, no Prism cloud.
- Explicit third-party requests — literature lookups and optional MinerU processing only fire when you start the action.
- Zero telemetry — no tracking, no analytics. See the Privacy Policy.
- Human veto — Plans, permission modes, and diffs keep the researcher in absolute control.
- Install — grab the latest
.dmg/.exe/.AppImagefrom Releases or the website.macOS Gatekeeper: if the app is reported damaged, run
xattr -cr /Applications/PrismNext.app. - Add a key — Settings → Models, pick a provider, paste your API key.
- Open a project — the workbench + opens a folder or scaffolds one from a template (Paper / Research Lab / Minimal). Your
.texstays in the repo root; metadata goes to.workbench/. - Pick a team and send — switch the active team to change how the agent works. Connect a lab machine anytime via Host → SSH.
Agent instructions live at .workbench/agent/AGENTS.md (Settings → Prompts & Rules). After creating or editing skills, open a new chat tab so the session picks them up.
Contributions to the open-source Host, research skills, and compiler integrations are welcome!
- Development setup: CONTRIBUTING.md
- License: Apache License 2.0 · Copyright © 2026 yibocat — see NOTICE
PrismNext — the full research loop, on your desk, under your gates.