See the whole network,
down to a single edge.
Atlas is a free desktop workbench for researchers and analysts who work with graph data of any size, from tens of nodes to tens of millions. State‑of‑the‑art algorithms, blazing‑fast performance, no server and nothing leaving the machine.
Free · macOS on Apple SiliconRead the user guide
- StyleMap colour, size, edges and labels to any property or analysis.
- LegendEvery class or community with its node count.
- TimelineDrag across it to draw only the edges active in a window.
- Status barSays what is drawn, what is aggregated and what is left out.

- SearchFind nodes by ID, prefix or any text field.
- InspectorEvery property and analysis result of what you select.
- CanvasThe whole graph on the GPU, from overview to single edges.
- StyleMap colour, size, edges and labels to any property or analysis.
- LegendEvery class or community with its node count.
- TimelineDrag across it to draw only the edges active in a window.
- Status barSays what is drawn, what is aggregated and what is left out.
- SearchFind nodes by ID, prefix or any text field.
- InspectorEvery property and analysis result of what you select.
- CanvasThe whole graph on the GPU, from overview to single edges.
The Atlas workspace on a 250,000-node graph.
Tens of millions of nodes.
Seconds, not hours.
Speed is the point. Atlas is a native Rust engine: bundles are memory-mapped, analyses run in parallel across your CPU cores, and layout and drawing run on the GPU. The size of the graph stops deciding what you can ask of it.
- 0.07s
- to reopen a 10-million-node graph. Bundles are memory-mapped, so size barely matters.
- 85s
- to import 40 million nodes and 93 million edges from raw tables
- 20s
- to find communities across all 40 million nodes with Leiden
- 43s
- to lay out every one of those nodes on the GPU
Benchmarks on a Mac Studio with Apple M2 Ultra and 64 GB of memory. Times depend on your hardware and your graph.
From the whole graph to the one edge that matters.
Atlas never hands you a sample. Zoomed out, dense regions draw as glowing groups with exact directed counts between them. Zoom in and they resolve into individual nodes and original edges, and the status bar always tells you what is aggregated.

Frames from atlas render on a synthetic graph of 3 million nodes and 10 million edges, Leiden communities, community-first layout.
One workflow, from tables to answers.
- Import
- Analyse
- Lay out
- Explore
- Ask
- Share
Bring your tables, keep their meaning
Point Atlas at CSV or JSONL node and edge tables, or node-link JSON from networkx or D3. Declare types, an edge type, a weight and a time role with its meaning: an event, an interval or a first observation. Error policies count every repair, and a rejected import publishes nothing.

Bundles are memory-mapped, checksummed and content-addressed: they open instantly and can be verified at any time.
Exact where it matters, honest where it can’t be.
A map, not evidence
Aggregates are exact directed counts, overview groups are labelled as spatial cells, and the status bar reports what is left out. Every layout carries the reminder that closeness on screen is not proof.
Direction first
Who links to whom matters. Bow-tie classes, condensation depth, directed (k,l)-cores, forward and reverse personalised PageRank, a directed triad census and density-corrected reciprocity.
Statistics with error bars
Degree-preserving null models with z-scores, likelihood-ratio verdicts on degree distributions, Leiden resolution ladders and stability across seeds.
Time as your data records it
Edge times keep their meaning, whether an event, an interval or a first observation, through time-respecting reachability, windows and difference maps.
Every view is replayable
Each change is recorded as Atlas Script, with undo, session logs and content fingerprints. A saved view warns you when the data under it has changed.
Colour for every analyst
Top-k colouring keeps categories apart under colour-vision deficiency, and the canvas can simulate protan, deutan and tritan vision.
Script it, render it, hand it over.
Everything you do in the app is a command. Record a session and replay it, drive the same catalogue from the command line, or reach the graph from Python without copying it.
# What feeds this node? The recorded answer, replayable.analyze ppr seeds=n4417 direction=reverse name=feedscolor analysis:feedspalette emberdoi in 2 300 n4417window 2025-02-01 2025-03-01saveview feeds-februaryEvery view change in the app is recorded in this language, with undo and replay.
Your graph stays on your machine.
Atlas is a native application on a Rust engine and the GPU. There is no server to stand up and nothing to upload: bundles live on your disk, analyses run on your CPU cores and drawing runs on Metal.
- Price
- Free
- Platform
- macOS on Apple Silicon
- Install
- Disk image · Homebrew
- Engine
- Rust · wgpu · Metal
- Scale
- Tens of millions of nodes
- Bundles
- Memory-mapped · checksummed
- Network
- None required
- Interfaces
- App · CLI · Script · Python
- Documentation
- User guide


