Hillclimb: Open-source AlphaEvolve-style search with Claude Code and Codex

Rebase Energy launched Hillclimb, an open-source command-line interface and modular Python framework that uses coding agents like Claude Code and Codex to search autonomously for Python programs optimizing a verifiable reward score. The project structures automated code discovery around problems verified by shell scripts and customizable hillclimbing search policies, according to github.com's report. By dispatches of headless coding assistants against programmatic scoring functions, the tool manages automated code optimization loops directly from the local terminal.

This report examines the framework architecture, execution commands, policy components, and setup workflow for software engineers, machine learning researchers, and performance optimization specialists. Readers will find the complete command set, catalog module options, Python SDK interfaces, and head-to-head benchmarking methods needed to deploy automated code synthesis pipelines without external dependencies beyond standard Python environments.

What shipped in Open-source AlphaEvolve-style search

Rebase Energy released Hillclimb under the MIT license as both a Python library requiring Python 3.12 or newer and a standalone CLI tool. The system organizes automated exploration into two decoupled primitives: the problem definition and the climber search algorithm. A problem specifies an evaluation standard using a verifier script named verifier.sh and an optional base solution named solution.py, while the climber controls solution drafting, selection, and mutation across iterative cycles.

The CLI ships with dedicated commands to manage search execution workflows. The run subcommand launches background exploration while releasing terminal control, allowing users to inspect live agent activity, progression charts, or decision trees using the watch subcommand. Search runs store optimal results locally under runs directories, while the summit command retrieves and saves the top-scoring solution directly to the workspace folder.

What improved in Open-source AlphaEvolve-style search

Hillclimb improves automated program search by shifting from monolithic prompt loops to a modular five-stage execution pipeline for each search step. The framework separates candidate selection, operator assignment, agent prompting, parameter tuning, and cross-search memory into independent, swappable components. Catalog climbers like openevolve and gepa can be downloaded directly into user workspaces alongside preconfigured greedy policies to facilitate comparative evaluation.

The search framework also adds head-to-head benchmarking capabilities through run spec flags and experiment reporting subcommands. Developers can test competing selector or operator policies against identical target problems, such as the included heilbronn-convex-13 challenge. The framework automates multi-seed parallel execution with configurable trial multipliers to compare disparate search strategies under uniform wall-clock and evaluation constraints.

What you gain from Open-source AlphaEvolve-style search

Developers gain a reproducible framework for autoresearch methods rather than a hardcoded coding assistant harness. By defining success criteria strictly through executable verification scripts or Python callables, teams ensure that generated code meets measurable performance benchmarks before adoption. The runner manages headless agent dispatches against configured compute budgets, preventing unconstrained token consumption or infinite loop failures during exploration.

The Python SDK enables programmatic composition alongside file-based CLI operation. Engineers can instantiate Problem and Climber objects directly, specify wall-clock time and maximum evaluation counts through Budget objects, and convert full search histories into structured data frames with the to_frame method. The system also supports step-by-step interactive debugging through start and step methods on Climber instances.

How to get Open-source AlphaEvolve-style search

Users can install Hillclimb using the uv tool manager or standard pip package installation on systems running Python 3.12 or newer:

uv tool install hillclimb

pip install hillclimb

To configure agent connectivity, initialize project folders, and fetch reference catalog components, run the following setup commands:

hillclimb connect claude

hillclimb init

hillclimb problem get heilbronn-convex-13

hillclimb climber get greedy

Optional tuning backends and catalog search policies require supplemental dependencies installed via pip:

pip install "hillclimb[optuna]"

pip install -r climbers/gepa/requirements.txt

Execution runs start with problem definitions and compute time limits, while manual evaluations score baseline solution files prior to launching automated agent searches:

hillclimb verify heilbronn-convex-13

hillclimb run heilbronn-convex-13 --budget 10m

What to watch after Open-source AlphaEvolve-style search

Following deployment, teams must monitor agent runtime requirements and external catalog dependencies when scaling search workloads. Connecting alternative coding agents such as Codex, Pi, or OpenRouter routes requests through different token accounting schemes, while running with the dummy agent flag allows developers to test pipeline wiring and verifier scripts without invoking language models or incurring API charges.

Resource limits specified through wall-clock flags and maximum evaluation counts govern background process termination. Users running comparative experiments across parallel search instances should inspect run logs and generated experiment reports to evaluate whether MapElites, greedy heuristics, or genetic programming operators yield superior optimization across algorithmic challenges, speed profiling, and accuracy benchmarks.