What is Mantis?
Mantis is a workbench for reading data as a map. You give it a corpus, it embeds every record, projects those embeddings down to two dimensions, and hands you a canvas where proximity means similarity. From there you select regions, inspect what is inside them, plot them, and ask an agent about them, all against the same layout. Mantis is built at the Kellis Lab, MIT CSAIL, and runs at mantis.csail.mit.edu.
The four words
Almost everything in this manual is built from four nouns.
| Word | What it is |
|---|---|
| Space | The container. Holds your data, its metadata, and one or more maps. |
| Map | One layout of that data. A space can carry several maps over the same records. |
| Point | One record: a row, a paragraph, a molecule, a cell, a file. |
| Bag | A named set of points you saved from a map, so a selection survives past the moment you made it. |
Core concepts
Cognitive cartography
A map is a picture of meaning, not of geography. Each point is one record; points that landed near each other were near each other in embedding space. Clusters, contours, and cluster labels are computed on top of that layout, so the shapes you see are structure in the corpus rather than decoration.
The two axes carry no independent meaning: reduction always targets 2 dimensions, and the 2.5D view mode is a camera angle, not a third data axis. Read neighborhoods, not coordinates.
Multi-modal embeddings
The same machinery works over more than text. Mantis ships creation paths for spreadsheets, plain documents, audio, video, web and academic search, GitHub repositories, SMILES molecules, and H5AD single-cell matrices. A space can carry several maps built from different columns, or different modalities, of the same records. See Source types for the full catalogue.
Agentic workflows
The Composer is the AI surface, and it is a panel in the workspace rather than a separate chat window. It runs in three modes: orchestrator (the full assistant, with map tools, workflows, and integrations), ask (read-only tools, no mutations), and visual tour (the same tools, answering as a sequence of steps that drive the map). Because the agent works through the same tools you do, a turn can change your selection, create a bag, or write a file into the shared workspace, and you watch it happen on your own map.
The default workspace
Open a space and the Default layout gives you six panels:
- Composer (top left): the agent chat, with slash commands, follow-up chips, and mode selection.
- Bags (below it): saved sets of points. New bag turns the current selection into one.
- Map (center): the canvas and its tool island. This panel cannot be closed.
- Tree: the cluster hierarchy, for navigating structure rather than pixels.
- Inspector: what is behind the current selection, one tab per selected point.
- File Browser: the workspace files shared between you and the agent.
That is six of roughly forty registered panels. The rest, including Reader, Summarizer, Keywords, Quicksheet, Plot Controls (Simple), Feature Analysis View, Journeys, Mantis Coding, User Logs, and Narrator, open from the View menu (Alt+V), which has a search box over every panel name.
The same menu carries the built-in layout presets: Default, Rich, Exploration, Quantitative Analysis, Coding, Document Search & Gathering, Text Reading & Analysis, Visual Brainstorming, and Full. Picking one replaces the current arrangement; picking the active one again clears the workspace.
Layouts are per browser
Panel arrangement is stored in the browser’s local storage, not on the server. The same space opened in two browsers can show two completely different layouts, and a panel that seems to have vanished is usually collapsed into a tabset overflow rather than broken.
What Mantis does not do
- It does not give you a third data dimension. Reduction targets two dimensions; 2.5D tilts the camera.
- It has no command palette. Quick commands live in the Composer’s slash menu.
- It does not make absolute position meaningful. Only relative position, distance, and cluster membership carry information.