Introducing scVIVA-Tools & scvi-tools MCP

Introduction

Alongside scvi-tools v1.5.0, we're shipping two new packages that round out the ecosystem in very different directions: scVIVA-Tools, which consolidates seven spatial transcriptomics models into a single toolkit, and scvi-tools MCP, which puts scvi-tools knowledge directly in reach of LLM coding assistants like Claude.

scVIVA-Tools: a unified toolkit for spatial transcriptomics

If you've ever had to juggle different APIs, installation instructions, and tutorial conventions for deconvolution, denoising, and niche analysis, scVIVA-Tools is meant to remove that friction.

Version 0.1.6, released alongside scvi-tools v1.5.0, bundles:

• scVIVA — a niche-aware VAE that jointly models a cell's own expression and its microenvironment, disentangling cell-intrinsic from environment-driven variation.
• ResolVI — corrects segmentation errors, background contamination, and cell-size bias in cellular-resolution data (Xenium, MERFISH, CosMx).
• DestVI — multi-resolution deconvolution of spot-based data (e.g., Visium) into continuous cell-type compositions.
• GimVI — joint imputation of missing genes between scRNA-seq and spatial datasets.
• Stereoscope — a two-stage generative model for spot deconvolution.
• Tangram — maps single cells onto spatial coordinates via optimal transport-style alignment.
• Harreman — new in v0.1.5, infers spatially-resolved metabolic gene programs and cell-cell metabolic / ligand-receptor communication using local autocorrelation and spatial proximity graphs.

Why a unified package?

Each of these models started life with its own API, its own docs, and its own release cadence — great for iterating quickly, but hard for users trying to build an end-to-end spatial analysis pipeline. scVIVA-Tools gives them:

• a single pip install,
• one consistent scviva namespace and AnnData-first API,
• a shared user guide, API reference, and 11 worked tutorials (one or more per model, plus preprocessing),
• shared infrastructure for GPU acceleration (via optional rapids extras) and SpatialData/squidpy interop.

Getting started

pip install scviva-tools
# or, with spatial I/O and GPU acceleration
pip install "scviva-tools[spatial,rapids]"

import scviva
import scanpy as sc

adata = sc.read_h5ad("my_xenium_data.h5ad")
scviva.model.ResolVI.setup_anndata(adata, layer="counts")
model = scviva.model.ResolVI(adata)
model.train()

adata.obsm["X_resolvi"] = model.get_latent_representation()

Swapping in scVIVA for niche-aware differential expression, or DestVI for Visium deconvolution, follows the same pattern — setup, train, extract. See the scviva-tools documentation for the full API reference and tutorials for each model.

What's next for scVIVA-Tools

scVIVA-Tools is under active development. Already in progress:

• DiagVI — a cross-modality model bridging spatial transcriptomics and spatial proteomics.
• AMICI — cell-cell interaction modeling.
• VIVS — spatial variable selection.
• NOLAN — niche detection: identifies and characterizes recurring cellular neighborhoods (niches) directly from spatial coordinates and expression, complementing scVIVA's niche-conditioned representation learning.
• Starfysh — histology-based deconvolution.
• SPARL — a vision-transformer-based foundation model for spatial proteomics imaging data, part of a broader push toward foundation models for spatial omics.
• CSDE (Corrected Spatial Differential Expression) — corrects for systematic errors like cell mis-segmentation and mislabeling in spatial transcriptomics before they propagate into false discoveries, by combining large automated annotations with a small manually-validated subset via prediction-powered inference (PPI) to recover unbiased estimates with valid confidence intervals.

Looking further out, expect standardized spatial benchmarking metrics (via scIB-metrics) and pre-trained reference models on public Xenium and Visium HD atlases.

scvi-tools MCP: scvi-tools knowledge for LLMs

The second release is a different kind of tool entirely: scvi-tools MCP, a Model Context Protocol server that gives LLM coding assistants — Claude Code, Claude Desktop, Cursor, Windsurf, OpenAI Codex CLI, or any other MCP-compatible client — structured, up-to-date access to scvi-tools knowledge: model documentation, tutorials, API reference, workflow templates, pretrained Hugging Face Hub models, and community FAQ.

It's a pure knowledge layer — no runtime model execution — so it's safe to point at any assistant you already use for coding. Under the hood it exposes 17 tools, from recommend_model (rank models by task and data type) and get_setup_anndata_guide (the exact setup_anndata() call for your data) to search_tutorials, get_hub_model, and a catch-all search_knowledge. All of it is baked into the package at build time, and refreshed monthly by CI against the live scvi-tools changelog, tutorials, GitHub issues, Discourse threads, and Hugging Face Hub registry — so recommendations stay current with new models like the ones described above.

Installation

pip install scvi-tools-mcp

Or skip the install entirely and let uvx fetch it on demand — this is what all the client configs below use.

Connecting to your LLM client

Claude Code (CLI)

claude mcp add scvi-tools-mcp -s user -- uvx scvi-tools-mcp

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "scvi-tools-mcp": {
      "command": "uvx",
      "args": ["scvi-tools-mcp"]
    }
  }
}

OpenAI Codex CLI — add to ~/.codex/config.toml:

[mcp_servers.scvi-tools-mcp]
command = "uvx"
args   = ["scvi-tools-mcp"]

Cursor / Windsurf / other MCP-compatible clients

{ "command": "uvx", "args": ["scvi-tools-mcp"] }

Once connected, ask your assistant something like "which scvi-tools model should I use for CITE-seq data with a reference atlas?" — it'll call recommend_model and walk you through setup, rather than guessing from a stale training cutoff. See the scvi-tools-mcp documentation for the full tool list and the repository for source and issues — and for a broader tour of LLM-assisted scvi-tools workflows (including the Claude skill bundle, ChatGPT, and other engines), see the user guide.

Try it out

The goal for both releases is the same: reduce the friction between having an idea and running the analysis, whether that friction is API fragmentation across spatial models or an LLM assistant guessing at how to call scvi-tools correctly. Please try them out, open issues, and let us know what you'd like to see next on GitHub or Discourse.