Ask Claude to "make this kernel faster" today and it guesses: it cannot see the Nsight report, does not know which kernel eats 60% of the step, has never read the SASS the compiler produced, and cannot benchmark its own rewrite on your card. hotkernel is the missing pair of hands. Ten tools wrap nsys, ncu, cuobjdump, torch.profiler and a strict verify-and-record contract; one skill turns them into a loop: profile, pick the hot kernel, read its SASS, check it against the roofline, rewrite it in Triton, verify against the baseline, record the verdict, keep or revert, repeat.

**Three targets, three autonomous agent runs, one RTX 5090: rmsnorm 3.2× faster than eager PyTorch, softmax 1.2× faster than torch.softmax, gemv level with cuBLAS. Every kernel the agent kept runs at 95–105% of the card's measured streaming bandwidth.**¹

¹ Each range comes from three independent verify() calls of the agent's final kernel against the baseline, n=7 interleaved runs each, on one RTX 5090 at 575 W under WSL2. Speedups are relative to the baseline scripts in examples/: eager PyTorch for rmsnorm, torch.softmax for softmax, torch.mv (cuBLAS) for gemv. "Measured streaming bandwidth" is device memcpy on the same card under the same timer, 1501 GB/s (84% of the 1792 GB/s spec), see the calibration. Raw logs for every number are in results/.

Install

Claude Code, one command (Claude Code 2.1.275 or later): it adds this repository as a plugin marketplace and installs the plugin, which is the MCP server plus the kernel-speedrun skill.

claude plugin install hotkernel --marketplace xinhuagu/hotkernel

Inside a session the same is /plugin marketplace add xinhuagu/hotkernel followed by /plugin install hotkernel@hotkernel. The plugin has one option, the Python that has your PyTorch and Triton: set it with /plugin configure hotkernel@hotkernel, or at install time with claude plugin install hotkernel@hotkernel --config hotkernel_python=$HOME/venvs/mytorch/bin/python. Left empty, hotkernel uses the first python on PATH that can import torch, and says so if there is none. The server itself stays torch-free and runs as uvx hotkernel.

Codex and other MCP clients (Cursor, DSH) keep the plain MCP setup:

codex mcp add --env HOTKERNEL_PYTHON=$HOME/venvs/mytorch/bin/python hotkernel -- uvx hotkernel

or {"command": "uvx", "args": ["hotkernel"], "env": {"HOTKERNEL_PYTHON": "..."}} in the client's MCP config, plus a copy of skills/kernel-speedrun/ in the agent's skills directory. Claude Code without the plugin works the same way: claude mcp add -e HOTKERNEL_PYTHON=... hotkernel -- uvx hotkernel.

Point it at your own kernel

hotkernel new turns any PyTorch function into a target with the same contract as the examples: the function becomes the baseline, its fp32 result the reference, tolerances default by dtype, and the protocol-3 timer is vendored next to the scripts, so they run in your torch interpreter whether or not hotkernel is installed there.

hotkernel new gelu --from torch.nn.functional:gelu --shape 4096x8192 --dtype bf16 --out targets

writes targets/gelu_baseline.py, targets/gelu_candidate.py (starts as the baseline; this is the file the loop edits) and targets/_hotkernel_contract.py. A function of your own works the same way, with extra inputs in call order: --from ./mykernels.py:fused --extra w:8192. Then bench the baseline for the number to beat, and verify the untouched candidate against it, which must come back inconclusive before anything changed. Agents get the same through the new_target tool. On the reference card the untouched gelu candidate benched at 0.0952 ms and verified inconclusive; a deliberately wrong one came back incorrect.

Before / after

All rows: RTX 5090 @ 575 W, bf16, timing protocol 3, three independent verify() runs of n=7 each.

² gemv is almost all reads; the copy ceiling is half reads and half writes, so a read-only kernel can sit slightly above it.

Every row on LEADERBOARD.md was written by record(), including the runs that did not win and a deliberately wrong candidate that verify() rejected as incorrect. Each agent run's full transcript is in results/<target>/agent-run-*/.

What the agent gets

torch_trace, bench, verify and record work on any machine with PyTorch. Nsight is optional and only makes the picture sharper.

The loop

The skill in skills/kernel-speedrun/SKILL.md is the whole method:

1. bench the baseline. That is the number to beat.
2. torch_trace or profile to find the kernel that actually dominates.
3. roofline with the kernel's flops and bytes: memory-bound kernels want fusion, compute-bound kernels want tensor cores and bigger tiles.
4. disasm the Triton kernel after its first compile: register spills, uncoalesced loads, long dependency chains are all visible in SASS.
5. One hypothesis, one change, then verify(candidate, baseline, n=7). It returns faster only when the gap beats 2 * sqrt(std_b²/n + std_c²/n) and every candidate run passed the correctness check.
6. record() the result, win or lose. It writes the log and the leaderboard; the agent never edits the board. Keep the change only if record() says it beats the run's best so far, otherwise revert.
7. Stop after three changes in a row that do not beat the best, or when the kernel is within 10% of the measured bandwidth floor.

It is autoresearch with kernel wall-clock as the metric instead of training loss.

Measured, not claimed

LLMs writing CUDA kernels have a credibility problem: the AI CUDA Engineer claimed 10 to 100× in 2025 and was found to be bypassing its own correctness checks, and KernelBench exists because "the kernel compiles" and "the kernel is correct and faster" are different things. hotkernel's answer is structural:

• The contract lives in the baseline. Shape, dtype, seed, reference, tolerances and the timer are in the baseline file the agent cannot edit; the candidate only supplies a function. verify returns invalid if anything in the contract differs.
• The timer was wrong twice and fixed in the open. Version 1 reused one input and let the 96 MB L2 pose as DRAM. Version 2 flushed L2 by zeroing a buffer and charged every call ~45 µs of someone else's write-backs; the softmax agent noticed that Nsight and verify disagreed by 2×. Version 3 rotates buffers larger than twice the L2 and was calibrated against five streaming kernels, which all land at 79–84% of the spec sheet, matching ncu's hardware counters. Old rows are kept and marked as not counted.
• Wrong is loud. verify reports no speedup for a run whose ok is false, every row carries n and std, and the baseline is in the repo so anyone can rerun it.

Platforms

• NVIDIA GPU required. Reference card: RTX 5090 (sm_120). The roofline table covers 5090, 4090, 3090 and H100, each with its source; add yours in src/hotkernel/gpus.py.
• WSL2 is the reference platform. All nine tools are verified on an RTX 5090 under WSL2 Ubuntu 24.04, driver 617.42: nsys 2026.1.3 and ncu 2026.2.1 extracted into user space from NVIDIA's apt repo (no sudo), cuobjdump and nvdisasm straight from the Triton wheel (no CUDA toolkit needed). ncu needs the GPU performance-counter permission once (NVIDIA Control Panel, Developer settings). Versions, commands and raw outputs: docs/platform-matrix.md.
• Native Windows: the server runs (pure Python). If your torch lives in WSL2, start the server through wsl.exe; the platform matrix has the exact claude mcp add line.
• Triton on RTX 50 series: FlashAttention 2/3/4 do not run on sm_120, which is exactly why hand-tuned Triton kernels matter on these cards.

License

Apache-2.0.