thermal-mcp-server

A Python package and MCP server that exposes simplified liquid-cooled accelerator thermal models as AI-callable tools for rack-level cooling analysis.

What It Models

• Steady-state 1D thermal-resistance networks for GPU cold plates.
• Coolant heat pickup from energy balance.
• Darcy-Weisbach pressure drop with simple laminar, transition, and turbulent handling.
• Water and 50/50 glycol comparisons using fixed nominal properties.
• Identical-GPU racks in series or parallel topology.
• First-pass CDU flow, pressure-drop, return-temperature, and junction-temperature sizing.
• Public accelerator reference cases with explicit source and estimate labels.

What It Does Not Model

• It is not a CFD solver or vendor thermal-design substitute.
• It is not validated against proprietary test data.
• It does not model manifold/header pressure losses, pump curves, fouling, transient/two-phase behavior, 2D spreading, detailed cold-plate geometry, or flow maldistribution.
• It does not support heterogeneous racks; each rack analysis assumes identical GPUs and cold plates.
• Financial and ROI modeling belongs outside this package.

Quickstart

git clone https://github.com/riccardovietri/thermal-mcp-server.git
cd thermal-mcp-server
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest
python examples/quickstart.py

Expected output from python examples/quickstart.py:

thermal-mcp-server quickstart

Single H100 SXM cold plate
  Heat load:        700 W
  Flow rate:        8.0 LPM water
  Inlet temp:       25.0 deg C
  Junction temp:    70.9 deg C
  Margin to 83 C:   12.1 deg C
  Pressure drop:    16.8 kPa
  Flow regime:      transitional

8-GPU rack, parallel topology
  Rack heat load:   5.6 kW
  Total CDU flow:   64.0 LPM
  Max junction:     70.9 deg C
  CDU return temp:  26.3 deg C
  Cold-plate dP:    16.8 kPa

For a series-vs-parallel rack comparison:

python examples/rack_sizing_example.py

MCP Usage

Install the package:

pip install thermal-mcp-server

Add the server to an MCP client:

{
  "mcpServers": {
    "thermal": {
      "command": "python",
      "args": ["-m", "thermal_mcp_server"]
    }
  }
}

If your MCP client does not inherit your shell PATH, use the absolute path to the Python executable inside the environment where thermal-mcp-server is installed.

Example MCP client run:

The client calls analyze_coldplate through the MCP server, receives the thermal result, and interprets the output in the conversation.

To reproduce this locally without a separate client, run the in-memory demo — it drives the same MCP server in-process (no network, no API key) and prints the exact request/response payloads a model sees:

python examples/mcp_client_demo.py

Tools

Tool	Purpose
analyze_coldplate	Single cold-plate thermal and hydraulic analysis
compare_coolants	Water vs 50/50 glycol comparison at identical conditions
optimize_flow_rate	Minimum flow search for a junction-temperature target
analyze_rack	Identical-GPU rack analysis in series or parallel topology
generate_decision_report	First-pass sizing memo with flow band, risk, uncertainty, and model blind spots

See docs/mcp.md for tool contracts.

Physics Assumptions

The model uses:

T_junction = T_inlet + 0.5 * coolant_rise + Q * R_total
R_total = R_jc + R_tim + R_base + R_conv
coolant_rise = Q / (m_dot * cp)

Heat transfer uses Dittus-Boelter for turbulent flow, Nu = 4.36 for laminar flow, and a linear blend in transition. Pressure drop uses Darcy-Weisbach with the same regime split.

Read the concise model notes in docs/model_overview.md and docs/assumptions.md. The detailed derivation and hand-calculation references are in docs/physics.md.

Public Reference Cases

The examples include H100 SXM, B200/NVL72-style, MI300X, and Gaudi 3 cases. Only H100 TDP and thermal limit are treated as vendor-published values in the default examples. Other limits, package resistances, and high-power cold-plate geometry are marked as estimates or proxies where vendors do not publish them.

See docs/public_specs.md and examples/real_chip_benchmarks.py.

Tests

The current suite has 75 tests:

• Physics behavior and hand-calculation checks.
• MCP wrapper contracts and error envelopes.
• Decision report behavior, including rack-aware feasibility.
• Smoke tests for examples/quickstart.py, examples/rack_sizing_example.py, and examples/mcp_client_demo.py.

Run:

pytest

Development

uv sync --group dev
uv run pytest
uv run ruff check .
uv run ruff format --check .
uv run mypy
uv build
uv run python examples/quickstart.py
uv run python examples/rack_sizing_example.py

These mirror the CI gate; all must pass before a PR can merge.

Roadmap

• Pump-curve support as an explicit input instead of a fixed pump-efficiency estimate.
• Transient thermal-capacitance model.
• More public reference cases with clearly labeled source quality.
• Interactive notebook polish for Colab use.
• Cold-plate optimization only as a separate experimental module or package.

Why This Exists

This project started as a way to explore how AI assistants can call lightweight engineering models directly instead of only producing static text. The package exposes simplified liquid-cooling calculations through Python and MCP tools, making it possible to ask design-tradeoff questions about accelerator cooling, rack-level CDU sizing, and coolant topology in a reproducible way.