BEFORE · TUTORIAL SOURCE
We trained LeNet-5 and a parameter-matched multilayer perceptron on a deterministic MNIST subset. Ten three-epoch CPU reference runs used five shared seeds per architecture.
✳ RESEARCH, IN THE AGE OF AGENTS
Plan experiments, follow your team, and write the paper together.
Your AI agent helps. Research Tab keeps the work connected.
FROM A REQUEST TO A RESEARCH RECORD
Three examples from one connected project.
Illustrative requests · Actual application captures · Fictional collaborators and tutorial measurements
PAPERS · COLLABORATIVE LATEX
Shared LaTeX, a compiled PDF, and the research behind it.
Ask for a revision or a reference. Review the suggestion.
Apply it to the manuscript your team is writing.
“Make the opening of this abstract clearer. Keep the methods and numbers intact.”YOU REVIEW & APPLY
BEFORE · TUTORIAL SOURCE
We trained LeNet-5 and a parameter-matched multilayer perceptron on a deterministic MNIST subset. Ten three-epoch CPU reference runs used five shared seeds per architecture.
REVIEWED REVISION · IN THE CAPTURE BELOW
We compared LeNet-5 with a parameter-matched multilayer perceptron on a deterministic MNIST subset. Each architecture was evaluated with five shared seeds in three-epoch CPU runs.
@article{lecun1998gradient,
author = {Yann LeCun and L\'eon Bottou and
Yoshua Bengio and Patrick Haffner},
title = {Gradient-Based Learning Applied
to Document Recognition},
journal = {Proceedings of the IEEE},
year = {1998}, volume = {86}, number = {11},
pages = {2278--2324}, doi = {10.1109/5.726791}
}SUGGESTED BIBTEX
Verify the metadata, add the entry to your bibliography, then cite it with \cite{lecun1998gradient}.
Agent assistance is illustrative. Researchers apply source changes; the current MCP tools read shared Paper source.
DISCUSSIONS, PEOPLE & PROGRESS
Ask what the team decided or how a collaborator’s work is going.
Get a concise answer grounded in shared records,
with the discussion and tasks there to inspect.
Progress reflects recorded task updates. Shared discussion context does not expose collaborators’ private local drafts.
EXPERIMENTS → FINDINGS → PAPERS
Compare runs, inspect the data, and carry figures into the paper.
Your agent can help surface a finding.
Researchers decide what the evidence supports.
Ten reference runs. Five seeds per architecture.
A preliminary interpretation, checked against the protocol.
Figures and tables retain their Result, Run and checksum references.
YOUR AGENT. YOUR WORKSPACE.
MCP is the connection that lets a compatible AI agent
read and update authorized research records.
You choose its access. You can revoke it.
Create a project connection with the permissions your agent needs.
Add the connection to an MCP-compatible client using the setup guide.
Give your agent a task, then review the resulting records in the dashboard.
Grant the permissions you need and revoke the connection when you choose.
Run work on your own machine. Choose what becomes part of the shared record.
Researchers review changes, verify evidence and make scientific decisions.
A place to start, with room for your research to grow. Invite collaborators through a research group; seat and project limits follow the plan shown below.
A complete individual workspace for adoption and genuine evaluation.
No. Your tools execute locally. Research Tab records the tasks, runs, metrics and artifacts you choose to share; it does not open a remote shell.
Yes. Papers supports collaborative LaTeX, compiled PDFs, review discussions and project evidence. You can export an Overleaf ZIP with the inserted research assets.
Current MCP tools can read shared Paper source. Ask for a paragraph revision or bibliography entry, then review and apply the suggestion in the editor.
Create or join a research group, then invite collaborators with the role they need. Your plan determines the available seats and active projects.
Eligible plans can connect institution-owned S3-compatible storage while preserving the same permission, scanning and provenance rules.
FROM THE FIRST QUESTION TO THE FINAL PAPER