Agents#
An AI agent can work in Cellucid with you: read the dataset shown, colour, filter, highlight, make annotation fields, run analyses, move through the history and explain what it finds. It acts only through the same actions the interface uses, so every change it makes is a step of your history, labelled with its name, and one click takes its whole turn back.
It can also do what lies outside the history: play and stop the guided tour,
make the printable report (as a file or as text for itself), export the
figure, the highlighted cells (CSV or .h5ad) or a field (CSV), list or
remove what this browser keeps (the Stored dialog) and forget a dataset
of My datasets, give Share’s public link of the dataset, and read and
vote on community annotation suggestions. None of these is a step, and
Undo does not take it back, so writing a file, removing what is kept and
publishing ask you first by default
(Permissions and what agents see). Voting publishes to the
repository, so it needs you to have connected the repository and signed in
to GitHub first: both need your click.
There are three ways in. All three use the same tools, the same permissions and the same rules for what is sent:
Way in |
What it is |
Read |
|---|---|---|
The chat bar (Ask an agent, at the foot of the plot) |
Cellucid’s own chat with Claude or OpenAI’s models, with your own API key, inside the tab |
|
The connector (Agents › Claude and ChatGPT) |
Claude or ChatGPT, in their own apps, reaching your open Cellucid tab through the Cellucid relay |
|
A local agent app ( |
Claude Code, Claude Desktop or Cursor on your computer, driving a viewer the Python or R package serves |
Local agents (MCP): Claude Code, Claude Desktop, Cursor, Local Agents (MCP): Claude Code, Claude Desktop, Cursor |
Whichever way an agent came in, what you see and decide is the same (Permissions and what agents see):
Your permissions decide what it may do without asking: changes, heavy work and what the history cannot take back are each Ask, Allow or Deny.
One line says which agent is acting, with Stop, and once it is done Undo turn.
What agents read goes to their AI provider. Once per dataset Cellucid says so, with the switch Agents see names and summaries only.
Text written by others is data. Names, labels, notes and a dataset’s own text reach an agent marked as data, never as instructions.
Important
The dataset’s files never leave your browser through an agent. What an agent reads (names, counts, summaries and, when it asks, values of cells or a picture of the plot) goes to the provider of its model with the conversation. See Permissions and what agents see and Privacy model.
Fast path#
You want to… |
Do this |
|---|---|
Ask a question about the dataset shown |
Ask an agent (or press |
Have Claude or ChatGPT work in Cellucid from their own app |
Agents › Claude and ChatGPT › Copy the connector address and add it in the app as a custom connector |
Use Claude Code or Cursor on a notebook’s viewer |
|
Take back what an agent did |
Undo turn, under its turn in the chat, on its line, or in the history |
Stop an agent now |
Stop on its line (or in the chat) |
Keep cell values and pictures from AI providers |
Switch on Agents see names and summaries only (the notice, or Permissions › Change…) |
Let it work without asking each time |
Permissions › Change…: a mode, or Allow per kind |