AIDE by Weco
AIDE from weco serves as an AI agent for Machine Learning. Its primary capability involves designing and optimizing Machine Learning pipeli...
Last verified:
What is AIDE by Weco?
Weco AI (Aide by Weco) is a platform for self-improving code that automates machine learning pipeline creation and code optimization. Powered by the AIDE algorithm—a Large Language Model-driven code optimization agent—Weco autonomously generates, tests, and iteratively refines candidate solutions against any metric you define, such as speed, accuracy, latency, or cost. The system keeps improvements and discards failed attempts, running continuous empirical exploration loops that can test hundreds of ideas overnight instead of requiring manual human iteration over weeks.
Key features include natural language interaction where users describe problems in plain English, automatic debugging and evaluation, iterative improvement based on previous experiments, and a visual dashboard showing a searchable tree of all candidate solutions. Weco works with any programming language (Python, C++, Rust, JavaScript) and any hardware stack (CPUs, GPUs), as long as the code prints a metric to stdout. It provides professionally crafted code and comprehensive research reports, supporting tabular data, time series, fraud detection, prompt engineering, inference optimization, and scientific ML tasks.
The tool targets machine learning engineers, data scientists, researchers, and even non-technical users who need to solve complex business challenges or academic research problems. It distinguishes itself from traditional AutoML by operating in the full code space rather than predefined modules, and from data science consultancy by automating weeks of manual heuristic work. Weco is built by frontier AI researchers whose AIDE algorithm achieved ~4× the medal rate of the next best autonomous agent across 75 Kaggle competitions on OpenAI's MLE-Bench.
AIDE by Weco pricing
Pricing model: Free
Free tier: 20 free credits (approximately 100 steps of Weco-hosted autoresearch) plus free access to Weco Observe dashboard for observability with your own autoresearch agent. BYOK (Bring Your Own Key) plan: Use your own OpenAI, Anthropic, or Gemini API keys with up to 10k experiments per month. Pay-as-you-go plan: Includes everything in Free plus access to all models and priority support. Installation: pipx install weco && weco setup claude-code.
AIDE by Weco pros
- Autonomously generates and tests candidate solutions without human intervention
- Optimizes against any custom metric: speed, accuracy, latency, or cost
- Works with any programming language: Python, C++, Rust, JavaScript
- Natural language interface makes it accessible to experts and novices
- Your data never leaves your machine—eval code runs locally
- Visual optimization tree shows every experiment in one searchable view
- Can run for weeks without human intervention for breakthrough results
- Each candidate costs fractions of a cent for extreme cost efficiency
- Delivers expert-grade production code plus comprehensive research reports
- Outperforms humans: ~4× medal rate on MLE-Bench Kaggle competitions
- Supports tabular data, time series, fraud detection, and scientific ML
- Side-by-side node comparison lets you inspect any two candidates
- Open-source CLI allows full transparency and verification
- steer optimization with natural language constraints like 'avoid unsafe memory'
- Integrates with Claude Code and Cursor via Weco skill
AIDE by Weco cons
- Currently alpha version primarily targets tabular data tasks solvable with CPUs
- Requires users to create their own evaluation script to measure performance
- May have comprehension limitations for complete novices without ML background
- No built-in hosting—you run locally or bring your own cloud infrastructure
- Limited multi-language documentation available for non-English speakers
- BYOK plan requires managing your own OpenAI/Anthropic/Gemini API keys
- GPU-intensive deep learning tasks still challenging compared to CPU tasks
- Learning curve for setting up evaluation scripts and optimization workflow
Frequently asked questions about AIDE by Weco
What is Weco AI and how does it work?
Weco is a platform for self-improving code that automates code optimization by iteratively refining your code against any metric you define. Powered by the AIDE algorithm, it systematically explores code variants, tests them empirically using your evaluation script, and delivers better-performing solutions. The system autonomously generates candidate solutions, keeps improvements, and discards failed attempts in a continuous loop.
What types of tasks can Weco optimize?
Weco excels at tabular data science tasks, time series analysis, fraud detection ML pipelines, prompt engineering for vision-language models, inference optimization for LLMs, and scientific ML for molecular property prediction. It can optimize Python, C++, Rust, and JavaScript code on CPUs or GPUs as long as your evaluation script prints a quantitative metric to stdout.
How do I get started with Weco?
Install via pipx install weco && weco setup claude-code, then ask Claude Code if it makes sense to apply /weco in your codebase. You need two things: a code file to optimize and an evaluation script that measures performance and prints a metric like 'speedup: 2.5x' to console output. The quickstart tutorial speeds up a simple PyTorch module in 10 iterations.
What is the AIDE algorithm?
AIDE is the LLM-powered code optimization agent that powers Weco. It achieved approximately 4× the medal rate of the next best autonomous agent across 75 Kaggle competitions on OpenAI's MLE-Bench, independently validated by researchers at OpenAI, Meta, and Sakana AI. AIDE generates expert-grade code by hypothesizing, experimenting, observing, and drawing conclusions following the scientific method.
Is my data secure with Weco?
Yes, your data never leaves your machine. The evaluation code runs locally where your data lives, and only metrics and code diffs are sent to Weco's servers. You can verify this by reviewing the open-source CLI code. This makes Weco suitable for sensitive data including transaction fraud data and proprietary business data.
What models does Weco support?
The Free tier uses Weco-hosted autoresearch. The BYOK plan lets you use your own OpenAI, Anthropic, or Gemini API keys. The Pay-as-you-go plan includes access to all models plus priority support. For the most economical setup, many users start with Gemini Flash-Lite's free tier of 1,000 requests/day.
How is Weco different from traditional AutoML?
Unlike conventional AutoML that operates within predefined modules or hyperparameters, Weco's design space is the entire code space, meaning it can perform any task a human data scientist can. It also provides structured knowledge in the form of code and research reports, not just outcome models. Users interact through natural language instead of learning complex APIs.
Can I see what Weco is doing during optimization?
Yes, each run produces a searchable tree of candidate solutions that you can inspect. Click any node to view the code, and compare any two nodes side-by-side. The Weco Observe dashboard provides web-based visualization and management of optimization tasks, with example runs available to explore.
How much does it cost to run experiments?
Each candidate costs fractions of a cent, finding non-obvious wins that manual iteration misses. The Free tier includes 20 credits (about 100 steps). With BYOK, you pay your API provider directly at cost price—most users generating moderate usage spend under $2/month on API costs plus the software subscription.
What if I'm using Claude Code or Cursor?
You can use the Weco skill for AI coding assistants. The skill guides your assistant through a structured workflow: analyze your code to infer the metric, establish a baseline, optimize using weco run, monitor progress, handle errors automatically, present results with clear explanations, and apply changes only with your permission.