@cluster2600/compilot-mcp

MCP server for cluster_compilot — an LLM agent proposes loop transformations, a polyhedral engine (ISL) proves legality, and clang -O3 +OpenMP compiles & times the result for a real speedup. This npm package is a thin launcher around the Python server; the heavy deps (islpy, clang) live in the Python environment.

Prerequisites

The Python package must be importable. Clone and install it once:

git clone https://github.com/cluster2600/cluster_compilot.git
cd cluster_compilot
python3.14 -m venv .venv && . .venv/bin/activate
pip install -e .
brew install libomp        # OpenMP for clang (macOS)

Then point the launcher at that interpreter via COMPILOT_PYTHON (or run it from the checkout, where it finds ./.venv/bin/python automatically).

Tools

| Tool | What it does | |---|---| | list_kernels | the schedulable kernels, by category (single / multi / stencil) | | check_legality(kernel, schedule) | prove a hand-written schedule legal (ISL) + compile & time it — single-statement, sub-second, no LLM | | optimize(kernel, backend=mock, iters, candidates, moa, aggregator) | run the agent loop, return best measured speedup + schedule. mock is offline/keyless; gemini/local drive real models |

Claude Code

COMPILOT_PYTHON=/path/to/cluster_compilot/.venv/bin/python \
  claude mcp add compilot -- npx -y @cluster2600/compilot-mcp

Codex

~/.codex/config.toml:

[mcp_servers.compilot]
command = "npx"
args = ["-y", "@cluster2600/compilot-mcp"]
env = { COMPILOT_PYTHON = "/path/to/cluster_compilot/.venv/bin/python" }

Try it

"List the compilot kernels, then check the legality of reorder(i, k, j) on gemm."

check_legality returns the ISL verdict and the measured speedup; optimize (default mock backend) runs the full search offline.