LLM agent orchestration
Corvex coordinates multiple LLM-powered agents that plan, call tools, and hand off subtasks, turning a fragile prompt chain into a repeatable scientific workflow.
Corvex — AI Agent Orchestration and Workflow Tooling
Corvex is an AI workflow tooling project that orchestrates LLM agents across scientific workflows. It ships an MCP server and client, routes models through LiteLLM, and runs closed-loop optimization with BoTorch and the Atomic Simulation Environment.
Overview
Scientific research workflows, especially in computational materials science, are long, repetitive, and scattered across command-line tools, calculators, and notebooks. Corvex wraps that pipeline in an agent layer. Instead of hand-running each step, an LLM agent plans the workflow, calls the tools it needs, parses results, and decides what to run next. The goal is to make AI tools a first-class part of the research loop, not an afterthought bolted onto a script.
Because the tool surface is exposed through the Model Context Protocol, the same workflow can be driven from any MCP-compatible client. The orchestration logic lives in Corvex; the front-end is interchangeable. That separation is what makes Corvex workflow tooling rather than a single-purpose app.
What it does
MCP server and client
A Model Context Protocol server exposes tools and data sources to any MCP-compatible client, so the same workflow plugs into different model front-ends without rewriting glue code.
LiteLLM model routing
Models are swappable behind a single interface. Route calls to the cheapest capable model for each step, or pin a specific provider for a sensitive stage, all from one configuration layer.
BoTorch Bayesian optimization
Closed-loop optimization runs through BoTorch, letting agents propose and evaluate experiments rather than sampling blindly, which converges faster on expensive simulations.
ASE atomic simulation
Integration with the Atomic Simulation Environment connects agents directly to atomistic calculators, automating the setup, execution, and parsing of materials science runs.
Streamlit interface
A lightweight Streamlit surface makes the orchestration layer explorable, so a researcher can launch a workflow, watch agent steps, and inspect intermediate results in a browser.
Tech stack
Why it matters
AI agent orchestration is most useful where workflows are expensive to run and easy to get wrong. By putting BoTorch optimization and ASE calculators behind agent-callable tools, Corvex lets a model propose the next experiment, run it, and learn from the result, automating the slow trial-and-error that dominates simulation-heavy research.