Frequently asked questions about Mantis

Mantis is cognitive infrastructure built at MIT CSAIL: it turns the data, knowledge, software, and workflows an organization already has into an environment where people and AI reason and act together. Records are embedded into a navigable map where proximity means similarity, nearby points are grouped into labelled clusters, and every conclusion stays traceable to its evidence.

Sources: What is Mantis?, About Mantis

Mantis is built at MIT CSAIL, in Professor Manolis Kellis’s group. The wider thesis extends that group’s approach of using computation to reveal hidden structure in complex systems, from genomics and biological circuitry to organizations themselves. The MIT and CSAIL names identify where Mantis is built; neither institution endorses the product.

Source: About Mantis

Clusters are auto-generated, labelled groups of nearby points. Mantis builds them with hierarchical clustering, so clusters nest and you can open one to read the finer structure inside it. Labels are written from the records a cluster contains.

Sources: Read the map, Clustering methods

A bag is a named, saved, coloured set of points. A selection is temporary and there is only one; you can keep as many bags as you like. Build one by filtering on fields, running a search, or hand-picking points on the map, then reuse it as the input to a plot, the scope for an agent, or the thing you hand to a colleague.

Source: Selections and bags

Mantis embeds each record with an embedding model, then reduces the result to two coordinates for layout and groups the points into a labelled cluster tree. The embedding, reduction, clustering and labelling configuration is generated from the backend’s own schema, so you can pick a hosted provider or an open model on the creation form, or supply vectors you computed yourself as a vector column.

Sources: Embeddings and projection, Source types

Access is set per space, with per-user permissions and separate visibility switches for the space and for each map. Vignette links are controlled on the same page. When Mantis is deployed inside your own environment, records stay in your storage, and open models can be served on your own hardware.

Sources: Sharing and visibility, Rooms and invites

Mantis is a web client backed by a Python service, a vector store and background workers. Maps are computed with standard embedding, dimension reduction and hierarchical clustering steps, so you can bring your own models and vectors. The workspace runs in the browser with nothing to install, and large maps stay interactive through a level-of-detail budget while you pan, zoom and filter.

Sources: Embeddings and projection, Source types

The Composer is a chat panel that reads your space, calls Mantis tools, and drives the map. It works through the same tools an analyst uses, its answers link to the points behind them, and Agent Studio lets you build a custom agent with its own knowledge and Model Context Protocol tools. The Mantis CLI gives any coding agent the same access from the terminal.

Sources: Composer, Agent Studio, Mantis CLI

Request access, sign in with Google or GitHub, and create a space from a spreadsheet, a folder of documents, a public source or a search. The tour walks one session end to end, from a first look at the map to a saved bag and an annotated region.

Sources: Get set up, Take the tour