ChatSpatial

[ChatSpatial Overview]

ChatSpatial replaces ad-hoc LLM code generation with schema-enforced orchestration. Instead of generating arbitrary scripts, the LLM selects tools and parameters from a curated registry, making spatial transcriptomics workflows more reproducible across sessions and clients.

ChatSpatial exposes 20 schema-validated MCP tools that orchestrate 67 spatial transcriptomics methods across 13 analytical categories. The tools are the stable natural-language interface; the methods are the analysis backends selected through tool parameters.

The server implements MCP 2026-07-28 through the official Python SDK v2 and continues to serve 2025-11-25 clients through SDK-managed protocol negotiation. STDIO remains the secure local default; Streamable HTTP is available for explicitly configured HTTP deployments.

Start Here

Install uv once, then register ChatSpatial without creating or managing a Python environment:

Codex:

codex mcp add chatspatial -- uvx --from chatspatial chatspatial server

Claude Code:

claude mcp add --scope user chatspatial -- \  uvx --from chatspatial chatspatial server

uvx creates an isolated environment on first launch and reuses its cache on later launches. Restart the MCP client after adding the server.

The command above installs the standard runtime. To make all 15 composable Python method families available in the same isolated MCP environment, use:

uvx --from 'chatspatial[full]' chatspatial server

full includes CellRank, FastCCC, the maintained spatial-domain and registration backends, annotation, enrichment, and the other portable Python families. R bridges, AESTETIK, and rctd-py remain separate because they have system, platform, or large-runtime requirements. See the installation guide before enabling those families.

Then:

1. Run your first analysis — Quick Start
2. Choose optional method families or a persistent environment — Installation Guide
3. Configure another MCP client — Configuration Guide
4. Reproduce the manuscript key results without an LLM — Reproduction image

Docker quick start:

docker pull cafferyyang777/chatspatial:latest

Minimal example prompt:

Load /absolute/path/to/spatial_data.h5ad and plot gene expression at the spatial coordinates

The Quick Start provides a downloadable sample and a Leiden workflow that runs with the standard installation. Its sample has spatial coordinates but no H&E image. SpaGCN requires the optional spatial-domains extra.

If you use Docker, mount host data to /data and prompt with the container path, for example /data/spatial_data.h5ad.

ChatSpatial works with any MCP-compatible client — Claude Code, Claude Desktop, Codex, OpenCode, and other MCP-capable tools.

Capabilities

Current coverage includes 67 methods across 13 analytical categories, exposed through 20 MCP tools. Supports 10x Visium, Xenium, Slide-seq v2, MERFISH, seqFISH.

Documentation

Reproducibility

Each release publishes a reproduction image under the -repro tag. It re-executes the DLPFC spatial-domain and OSCC CARD deconvolution tool calls recorded during the manuscript experiments, without calling any LLM, and compares the results with the reported values:

docker run --rm -v "$PWD/out:/outputs" \  cafferyyang777/chatspatial@sha256:ac3ba1ed8235c3d53ca72005d8ad2556cac007f5ddb932ae3e1d1241fa2cf9d8 \  chatspatial-reproduce

reproducibility/ contains the image definition and describes what it checks. The experiment and analysis scripts for the manuscript are in ChatSpatial-Reproducibility.

Citation

If you use ChatSpatial in your research, please cite:

@article{Yang2026.02.26.708361,  author = {Yang, Chen and Zhang, Xianyang and Chen, Jun},  title = {ChatSpatial: Schema-Enforced Agentic Orchestration for Reproducible and Cross-Platform Spatial Transcriptomics},  elocation-id = {2026.02.26.708361},  year = {2026},  doi = {10.64898/2026.02.26.708361},  publisher = {Cold Spring Harbor Laboratory},  URL = {https://www.biorxiv.org/content/early/2026/03/01/2026.02.26.708361},  journal = {bioRxiv}}

ChatSpatial orchestrates many excellent third-party methods. Please also cite the original tools your analysis used.

Contributing

Documentation improvements, bug reports, and new analysis methods are all welcome. See CONTRIBUTING.md.