Weco optimizes code against an evaluation metric you define. You point it at one or more source files and an evaluation command that prints the metric, and it runs an LLM-guided tree search: it applies a change, runs the evaluation, parses the result, and proposes the next change based on the goal. The search algorithm is AIDE, and results improve with more steps, so research-level tasks can take many hours of exploration.
Typical targets are agent harnesses (prompts, tools, and scaffolding, scored on accuracy, win rate, or cost), prompts (refined or compressed while holding accuracy), model development (feature transformations, architectures, or a whole training pipeline), and general performance work such as CUDA or Triton GPU kernels scored on latency, throughput, or memory bandwidth.
Weco also ships as a skill for AI coding assistants (Claude Code, Cursor, Codex, and OpenClaw): you describe what to optimize, and the assistant writes the evaluation, configures the run, monitors iterations, and explains the results. Runs appear in a hosted dashboard with tree visualization, code diffs, and metric tracking, and the observe mode logs experiments from your own optimization loop to the same dashboard without handing over control. Usage is billed in Weco credits, or with your own provider API keys.
Features
- Metric-guided tree search: explores and refines code variants with the AIDE algorithm, keeping the best solution for the goal you set
- Any measurable objective: the evaluation command prints the metric, so accuracy, win rate, cost, latency, AUC, or a custom score all work
- Coding-assistant skill: installs into Claude Code, Cursor, Codex, and OpenClaw so the assistant sets up and runs optimizations from a plain-language request
- Observe mode: tracks experiments from your own agent loop or script in the dashboard, with a CLI and a Python SDK
- Run management: JSON subcommands for status, ranked results, per-step code, diffs, mid-run instructions, and graceful stops
- Review mode: inspect pending nodes, replace their code, and submit them for evaluation
- Resumable runs: continues an interrupted run from its most recent node with all progress and metrics preserved
- Model choice: OpenAI GPT and o-series, Anthropic Claude through Vertex AI, Google Gemini, hosted open models through the Concentrate AI gateway, and Kimi through the Moonshot API, with automatic selection when none is set
- Bring your own key: provider API keys can be passed per run instead of using Weco credits
