clusters-mcp-server

Galaxy Cluster Catalogs MCP Server. An MCP server for public SZ and X-ray galaxy-cluster catalogues:

• Planck PSZ2

• SPT-SZ 2500d (Bocquet+19), SPTpol 100d, SPT-ECS and SPT-3G Main 5yr

• ACT DR6

• eROSITA eRASS1, primary and cosmology samples

Each catalogue downloads on demand and is harmonised to one schema.

Tools. The server includes tools to:

• plot the multi-survey M500–z plane and Mollweide sky maps

• build N(z) and N(>M) counts

• crossmatch catalogues by angle and redshift, and compare mass calibrations

• predict LCDM cluster counts with a Tinker+08 HMF × Eisenstein–Hu P(k), pure numpy/scipy, dispatchable to HPC

It is part of the cosmo-*-mcp-server family (port 8014), built from cosmo-mcp-template. See docs/CONVENTIONS.md.

Datasets

• All URLs were verified by download on 2026-11-04.

• The eROSITA bare .fits URLs return HTTP 403; the .fits.tgz archives work.

• The cache lives in ~/.cache/cosmo-mcp/clusters-mcp-server/; override it with $COSMO_MCP_DATA.

Harmonised schema. Every catalogue is written with these columns:

• name, ra, dec [deg]

• z, z_err

• M500 [1e14 Msun, M500c], M500_err_lo, M500_err_hi

• snr: the survey-native statistic

• survey, dataset

Masses come from each survey's own scaling relation. Run describe_catalog_columns to see which. In particular, Planck PSZ2 MSZ is hydrostatic and sits about 20% low.

Related MCP server: esa-gaia

Tools

The HMF kernel (tools/lcdm/kernels.py, lcdm.kernels.predict_counts) needs only numpy and scipy:

• Accuracy: agrees with colossus Tinker08 + eisenstein98_zb to 1–2% for 3e13–1e15 M☉/h, and to ≤5% at 3e15.

• Speed: about 0.2 s per prediction.

Skills

• clusters-tour

• multi-survey-mass-redshift

• sz-xray-crossmatch

• cluster-counts-vs-lcdm

• hpc-dispatch-handoff

Key numbers (tour, verified)

• ACT DR6 × PSZ2 (3′): 386 matches. M_Planck/M_ACT = 0.80 ± 0.02, ln-scatter 0.34. This is the Planck hydrostatic bias, 1−b ≈ 0.8.

• SPT-SZ × eRASS1 (2′): 330 matches. M_eRASS1/M_SPT = 0.95 ± 0.01.

• SPT-SZ, M500 > 5e14, z > 0.25: LCDM (σ8 = 0.8, Ωm = 0.3) predicts 138 against 105 observed. The Poisson-only best fit is σ8 ≈ 0.78 (Bocquet+19: 0.77).

• SPT-3G, M500 > 3e14: 542 predicted, 550 observed.

Install / run

uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -e ".[dev]"
MPLBACKEND=Agg .venv/bin/python -m pytest -q        # includes network tests
.venv/bin/python tests/smoke_server.py
PYTHONPATH=. MPLBACKEND=Agg .venv/bin/python scripts/run_tour.py output/tour
./run_server.sh                                     # stdio; or: .venv/bin/python -m mcp_server --transport http --port 8014

Caveats

• predict_cluster_counts_lcdm has no selection function, mass–observable scatter or Eddington bias. It is a sanity check, not a cosmology constraint.

• Cross-survey masses use different calibrations, with 5–25% offsets between them.

• PSZ2 and eRASS1 count deficits relative to LCDM reflect incompleteness and the mass scale; they do not indicate new physics.