What it does
Weco improves code using the results of your own evaluation. You point it at one or more source files, give it a command that prints a metric, say whether to maximize or minimize that metric, and set how many steps to run. A tree search guided by language models then tries changes, runs your evaluation, reads the result and proposes the next improvement. The README names several kinds of work: refining prompts, improving the scaffolding around LLM agents, tuning model architectures or whole training pipelines, and speeding up code such as CUDA or Triton GPU kernels. Interrupted runs can be resumed, run status, results and diffs come back as JSON for scripts and coding agents, and observe commands log experiments from your own optimization loop to a dashboard. Weco also installs as a skill for Claude Code, Cursor, Codex and OpenClaw, so an assistant can set up and monitor runs from a plain-language request. You can pick from a long list of models from OpenAI, Anthropic, Google and others, or supply your own provider API keys. Runs use Weco credits, which you can check and top up from the CLI.
Capabilities
- Optimizes one or several source files against a metric printed by your eval command
- Tree search guided by LLMs, with resume for interrupted runs
- JSON output for run status, results and diffs
- Skills for Claude Code, Cursor, Codex and OpenClaw
- Observe commands that log your own experiments to a dashboard
- Models from several providers, or your own API keys
When to use it
- You can measure success with a script and want an LLM to iterate on prompts, models or kernels
- You want an AI coding assistant to run optimization loops for you
- You run your own experiments and want them tracked in one dashboard
