cxg-census-mcp

An MCP server that lets LLM agents query the CZ CELLxGENE Discover Census single-cell atlas without lying about it — ontology-aware filters, cost caps, full provenance + attribution on every response. Drop it into Cursor / Claude Desktop / Claude Code and ask questions like "Visualize the cell-type composition of the human lung" in plain English.

Independent / unaffiliated. Not affiliated with, endorsed by, or sponsored by the Chan Zuckerberg Initiative (CZI), EMBL-EBI, the U.S. Census Bureau, or anyone else. "CELLxGENE" is a CZI mark; references here are descriptive (nominative) use only.

No warranty. MIT-licensed source, "as is". Research/exploration tool — not a clinical or diagnostic instrument. Always verify results before publication. See LICENSE for the full trademark and content attribution notice, and SECURITY.md for the threat model and known-issues policy.

Alpha (v0.1.2). CHANGELOG.md

Demos

Healthy vs COVID-19 lung, side-by-side. Two parallel queries, the disease_multi_value_v7 schema-drift rewrite kicks in for the COVID cohort, attribution from both contributing dataset sets surfaces in the same chat turn.

https://github.com/user-attachments/assets/c836f225-5075-4643-87aa-70d311bc5fd2

Cell-type composition of human lung in one query. Free-text "lung" resolved to UBERON:0002048, routed through tissue_general, every CURIE labeled, all in a single Tier-0 call.

https://github.com/user-attachments/assets/b0e10ca7-e46b-4e5f-ae63-11949d328c4d

(Videos render on GitHub. On PyPI they appear as bare URLs — head to the GitHub README to watch.)

More prompts in docs/example-questions.md.

Architecture at a glance

                 ┌──────────────────────────────────────────────┐
   MCP client    │   tools/        thin MCP wrappers, no logic  │
   (Claude,  ─►  │     │                                        │
    Cursor,      │     ▼                                        │
    Code, …)     │   planner/      FilterSpec → QueryPlan,      │
                 │     │           cost estimate, tier routing  │
                 │     ▼                                        │
                 │   ontology/     OLS4 + hint overlay,         │
                 │     │           CL/UBERON/MONDO expansion    │
                 │     ▼                                        │
                 │   execution/    Tier 0  facet counts         │
                 │     │           Tier 1  chunked obs scan     │
                 │     │           Tier 2  expression aggregate │
                 │     │           Tier 9  refuse → snippet     │
                 │     ▼                                        │
                 │   clients/      OLS4 (HTTPS) + Census/SOMA   │
                 │                                              │
                 │   caches/       OLS, facet, plan, filter LRU │
                 │   models/       Response envelope w/         │
                 │                 attribution + provenance     │
                 └──────────────────────────────────────────────┘
                                    │
                                    ▼
                       ┌────────────────────────┐
                       │ EBI OLS4 (ontology)    │
                       │ CZ CELLxGENE Census    │
                       │ (CC BY 4.0 data)       │
                       └────────────────────────┘

Full architecture notes: docs/architecture.md. Tool reference: docs/tool-reference.md. Example questions: docs/example-questions.md.

Install

From PyPI (recommended):

uv tool install "cxg-census-mcp[census]"
cxg-census-mcp                       # speaks MCP over stdio

Or with pip:

pip install "cxg-census-mcp[census]"

Without the [census] extra you get mock mode (deterministic fixtures) — handy for offline demos and verifying your MCP client config without pulling tiledbsoma's ~1 GB of native deps.

From source (for development):

git clone https://github.com/MaxMLang/cxg-census-mcp
cd cxg-census-mcp
uv sync --extra dev --extra census
uv run cxg-census-mcp

MCP client config

Cursor (~/.cursor/mcp.json) and Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS) both expect the same shape. Cleanest is uvx once installed from PyPI:

{
  "mcpServers": {
    "cxg-census": {
      "command": "/absolute/path/to/uvx",
      "args": ["--from", "cxg-census-mcp[census]", "cxg-census-mcp"]
    }
  }
}

Use the absolute path to uvx (which uvx from your shell). MCP clients spawn the server in a non-interactive subprocess that doesn't source your shell rc, so a bare "uvx" will fail with No such file or directory.

If you cloned from source instead, point at the checkout:

{
  "mcpServers": {
    "cxg-census": {
      "command": "/absolute/path/to/uv",
      "args": ["--directory", "/path/to/cxg-census-mcp", "run", "cxg-census-mcp"]
    }
  }
}

Claude Code:

claude mcp add cxg-census -- /absolute/path/to/uvx --from "cxg-census-mcp[census]" cxg-census-mcp

Quit + relaunch your client (⌘Q on macOS — closing the window isn't enough) and the server should show up in the MCP panel with 13 tools.

Tools (13 total)

Workflow: census_summary, get_census_versions, count_cells, list_datasets, gene_coverage, aggregate_expression, preview_obs, export_snippet, get_server_limits.

Inspection: resolve_term, expand_term, term_definition, list_available_values.

Plus MCP resources (markdown docs at cxg-census-mcp://docs/{slug}), prompts (census_workflow, disambiguation), and cooperative progress / cancellation notifications. Details in docs/tool-reference.md.

Configuration

All env vars use the CXG_CENSUS_MCP_ prefix. Most useful:

Full list and validation: src/cxg_census_mcp/config.py.

Development & operations

Quick loop:

make install-all                 # uv sync --extra dev --extra census
make lint typecheck test         # ruff + mypy + pytest (mock mode)
make cov                         # tests + coverage HTML in ./htmlcov
make audit                       # pip-audit on locked production deps

Operational tasks (cache pre-warm, schema diff, container build, metrics dump, plan-cache vacuum, weekly hint/facet refresh) live in the Makefile and are documented in docs/operational-playbook.md.

Documentation index