Roadmap
This page describes where the project is heading. Items here are directions, not commitments — the Quick Start and Architecture pages describe what the server does today (DuckDB SQL over Parquet on S3, H3 spatial indexing, STAC-driven discovery). The roadmap is the bridge from that proven core toward the broader vision.
Beyond tables: array data via Zarr
Today the server queries tabular cloud-native data (Parquet) through DuckDB. A growing share of science data is array data — gridded rasters, imaging stacks, multidimensional cubes — for which Zarr and xarray are the cloud-native standard. We aim to extend the server so an agent can discover and query Zarr arrays through the same STAC-grounded interface it uses for tables, without leaving the cloud-native path.
This matters most for the life sciences, where spatially-resolved assays — spatial transcriptomics, whole-slide microscopy, 3D-genome imaging — are natively array-shaped and arrive at terabyte scale. Conventions like OME-Zarr and AnnData already point the way; the missing piece is the agentic layer that makes them reachable.
Hardware-accelerated engines (Polars / GPU)
The server confines agents to validated cloud-native engines. DuckDB on CPU is the engine today. A working GPU-accelerated prototype — mcp-gpu-data-server, built on Polars / cuDF with KvikIO for fast S3 reads and partition pruning — already exists and is being benchmarked head-to-head against this CPU server (tracked in issue #42). Early findings are nuanced: S3-I/O-bound queries favor CPU, while compute-heavy queries on already-loaded data favor GPU. The goal is to let an agent transparently use whichever engine fits the query, all behind the same MCP interface.
Broader domains
The architecture is domain-agnostic — any data with a tabular or array structure and an indexable dimension fits the same pattern. Proven first at earth-observation scale, the work is being extended to the life sciences in collaboration with a lab that co-developed image-based spatial transcriptomics. As the vision describes, the aim is a single agentic interface to the cloud-native stack across scientific domains.
Smaller, more open models
The server already injects all query guidance at call time, so compact open models can drive it. We are continuing to tune that guidance so increasingly small, locally-run open models can perform real analyses reliably — keeping sensitive data on local hardware and reducing dependence on closed models.
Have a use case that needs one of these sooner? Open an issue — concrete use cases drive prioritization.