Overview

Weco is an autoresearch engine for code. You point it at your codebase, give it a way to measure success, and it runs the experiments for you - proposing a change, running your evaluation, keeping what improves the metric, and building on it - until it finds the best-performing version.

How it works

Weco uses LLMs and tree search (our AIDE algorithm) to optimize code against a metric you define. You give it code to optimize and an evaluation script that prints a number (for example, speedup: 2.5). Weco explores a tree of code variants - each node a change it proposed, ran, and measured - and when it finishes, your file holds the best version it found.

It works with any language and on any hardware - the only requirement is that you can express "better" as a number.

Why Weco

Weco doesn't just provide the optimization loop, it gives you control over the research process itself:

• Traceable. Every experiment is a node in a solution tree you can explore in the dashboard - see what was tried, what each variant scored, and exactly how the best solution was reached. No black-box answer you have to take on faith.
• Steerable. When a run plateaus or you have an idea - from a paper, a hunch, domain knowledge - branch the search mid-run with a natural-language instruction instead of starting over. Fan out parallel branches to test competing hypotheses side by side.
• Yours. Code and data stay on your hardware; Weco orchestrates the search, your machine runs the evaluations. You keep full IP and can bring your own model keys.
• Agent-native. Drive it from Claude Code, Codex or Cursor with the Weco skill.

When to use it

Weco is a good fit when:

• You have a measurable goal - speed, accuracy, latency, memory, cost, throughput, or a quality score.
• The problem rewards trial and error - there are many ways to write the code and the best one isn't obvious up front.
• You can write (or already have) an evaluation script that runs the code and prints the metric.

Common uses include optimizing GPU kernels (CUDA, Triton), tuning ML models, engineering prompts, and speeding up or shrinking general-purpose code.

It's not the right tool when there's no way to measure success as a number, or when a single deterministic edit would do. Weco's value comes from iterating against a metric.

Get started

The fastest way to start is to let your AI coding assistant install and set up Weco for you. Visit our Quickstart to get optimizing.

If you'd rather drive the CLI yourself, see the Manual Installation instructions and Optimize with the CLI.