Voxel51 Launches 3.0 Multimodal Data Platform With Built-In Agent

The company describes three core additions. Time-series multimodal support lets teams view multiple sensor streams on a single timeline, with segment-level embeddings used to locate specific moments inside long recordings. Unified annotation covers 2D, 3D, video and multimodal data, and includes automated quality assurance, annotator metrics and human-in-the-loop review. Agentic Labeling uses vision-language models to generate labels from natural-language instructions and examples. Supported modalities include MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial data and audio.

The new Voxel51 Agent ships with more than 45 MCP tools, more than 80 operators and 18 open-source skills. It chains search, curation, annotation and evaluation into repeatable workflows and can write code to build custom plugins. Voxel51 cites a robotics team searching behaviors across thousands of episodes to curate balanced training sets, and a perception team starting from a model failure and pulling similar long-tail scenarios for labeling. All workflows inherit the organization's permissions and governance controls.

Voxel51 Search is a persistent hybrid index that indexes datasets and computes embeddings automatically. Semantic retrieval finds visually or semantically similar moments even without shared labels, while lexical and structured queries handle exact terms, timestamps, sensor fields and metadata. Embeddings and structured fields are kept in native form and combined at retrieval time; vector retrieval runs on LanceDB with vector, scalar and full-text indexes, separating storage from compute. Incremental updates avoid full rebuilds when samples arrive or schemas change.

The release also adds four provenance layers: a deployment-level activity log recording who did what and when, a dataset-level log of how a single dataset evolved, per-sample edit history for label, metadata and property changes, and dataset versioning through immutable snapshots for reproducing training and evaluation runs. Events flow through fiftyone-mq for job scheduling and Activity Core for event tracking.

The platform runs as a hosted web app at app.voxel51.com or self-hosted in cloud, hybrid, on-premises and air-gapped setups. No local GPU requirement was disclosed, and no product benchmarks were published; Voxel51 instead cites its own survey data. The free tier includes Voxel51 Agent, 30,000 inference requests and 30 compute hours per month, capped at three datasets of 10,000 samples each and 20 GB of storage. The Team tier offers eight seats and 2,800 compute hours per month at a price the company did not disclose. The platform is model-agnostic across embeddings, auto-labeling, inference and evaluation, so a team can start with a foundation model and later swap in its own domain model.

Voxel51 points to its 2026 State of Physical AI survey, which found that 58% of teams it classifies as exceptional spend more than half their project time on data work, compared with 21% of struggling teams. The research was run by the vendor.

In comparisons reported by MarkTechPost, Encord does not disclose its free tier, while Foxglove offers 10 GB of storage, three users and five devices free, with a Pro plan starting at $20 per month plus usage and 1 TB of storage. Foxglove's MCP support is a local MCP on paid plans; Encord's sensor-stream support was stated in its Series C release. Prices were verified on October 17, 2026.