# The open source toolchain for multimodal data

Use the Rerun SDK to log, store, query, view, and train on physical data. Flexible high-performance building blocks for your data loop.

![A synthetic urban LiDAR scan with vehicles, pedestrians, and roadside structures.](/scenes/nuscenes.svg)

## Visualize everything

See exactly what your robot saw: synchronized cameras, joint states, and 3D in motion. Debug episodes, review datasets, and compare teleop with policy rollouts.

## Query and transform

Analyze and reshape your whole dataset. Derive progress scores, action labels, or normalized schemas, then add the results as new layers alongside your source data.

## Log, store, and convert

Log straight from your code or convert existing data from MCAP, ROS 2, or Parquet. One format carries your data through visualization, queries, and training.

## Train

Train on the same data you record and query. A dataset mix is just a query, fed straight to your model, with no separate training set format to maintain.

- [Read the docs](/docs)
- [View on GitHub](https://github.com/rerun-io/rerun)

## Inside the SDK

Every part of the Rerun SDK is open source and built on the same data model, so the pieces fit together. That includes APIs and a `rerun` CLI for spinning up servers and viewers, inspecting recordings, and converting files.

### Logging library

Log physical data with idiomatic Python, Rust, and C++ APIs. A shared semantic model keeps sensor readings, model outputs, and system state in context.

### Column-chunk file format

The .rrd format is designed around column-chunks, and is purpose-built to enable flexible and performant processing of multi-rate, multimodal data.

### Query and processing APIs

Query recordings as dataframes or with SQL. Slice the columns, time ranges, and frames you need, and reshape data with declarative chunk processing.

### Performant viewer

Build custom views with a fast, extensible Viewer. Run on desktop, in notebooks, or in the browser, and render headlessly for agents and automated jobs.

### PyTorch dataloader

Train directly on your .rrd files. A column and video-codec aware dataloader that streams dataset mixes to training without needing to re-export.

### Catalog server

The open-source catalog that indexes a folder of recordings, so finding and streaming across many files is a query. The local foundation the whole stack runs on.

## Built for engineers. Ready for agents.

**Let agents use the Viewer** through MCP to inspect data and control a running session. Give them a visual workspace for investigating recordings alongside you.

**Give agents structure** with entities, components, and timelines. Dataframes and SQL let them query the data they need instead of sorting through files.

**Render evidence headlessly** for agent workflows and automated jobs. Generate visual output from your recordings without opening an interactive Viewer session.

**Keep every action inspectable** in the Viewer. Review what an agent selected or generated alongside the source data, so you can check its work in context.

## One toolchain. Two ways to work.

Feature | Rerun SDK | Rerun Hub
--- | --- | ---
Scalability | **Flexible building blocks** for local workflows and integration with your own infrastructure. | **Managed infrastructure** for petabyte-scale datasets, with automatic scaling.
Data storage | **Portable `.rrd` files** and an open-source catalog for organizing and querying recordings. | **Persistent, managed catalog** with byte-range indexing across your own S3-compatible object stores. Data stays in your buckets.
Performance | **Fast local iteration** with a performant Viewer, query APIs, and direct access to your recordings. | **Fast queries at scale** with cloud-based indexing, using the same SDK and Viewer.
Collaboration | **Share recordings** through your existing team workflows, with the same open-source tools for everyone. | **Team-wide access** to the same recordings via authenticated share links. Everyone works from the same data.
Access control | **Your environment, your policies.** Manage access through the machines and infrastructure where you run the SDK. | **SSO authentication**, authenticated share links, and single-tenant isolation deployed in the cloud region you choose.
Compliance | **Audited SDK and Viewer**, with deployment and compliance policies managed by your team. | **Managed compliance** with SOC 2 and ISO 27001 certifications, 24/7 monitoring, and continuous audits.
Support | **Community support** via Discord and GitHub. Open source under Apache-2.0 / MIT. | **Commercial support** from a shared Slack channel to a dedicated forward-deployed engineer.

## Start in two minutes

Install rerun-sdk. The Viewer, the format, and the toolchain are open source. Use it for research, production, and everything in between.

- [Read the docs](/docs/getting-started/quick-start)
- [View on GitHub](https://github.com/rerun-io/rerun)