Download
Kapso is a self-improving software factory for AI/ML objectives. Give it an objective and it autonomously builds the best solution, deploys it, and learns from every run.
Kapso compounds. It learns from its own work, keeping what moved the metric and dropping what did not, each lesson backed by the evidence that earned it. And it learns from the public state of the art, so the field's best practices are always at hand.
Optimize CUDA kernels for GPU performance. Cut a PyTorch model's wall-clock and memory without changing the numerics. Deliver a prediction model end to end, from data prep to a deployable artifact. Evolve the harness around a model: prompts, decoding, scoring. Set agents to improving agents until the metric climbs. Works with Claude Code and Codex.
Proven in public: #1 open-source on MLE-Bench, #1 on ALE-Bench, ahead of tabularFMs on RelBench, above every human at IOAI 2026.
Docs: docs.leeroo.com/docs.
Source: github.com/Leeroo-AI/kapso.
Built by Leeroo.
Features
- Runs campaigns toward a measurable AI/ML objective
- Learns from every run: keeps what moved the metric, backed by evidence
- Learns from repositories and papers into a searchable knowledge graph
- Works with Claude Code and Codex
- Ships the winner: local, Docker or Modal deployment
- Benchmarked in public: MLE-Bench, ALE-Bench, RelBench, IOAI 2026
