TeXRA: Multi-agent AI for theorists

Featured
Personal - Present
LLM agentsTypeScriptLean 4WolframVS Code

A team, not a chatbot. An orchestrator decomposes a research task and delegates to specialist agents (search, derivation, numerics, review, Lean 4 formalizers), each with its own tools and model. Every claim is grounded in real databases and compiled figures, and every edit lands as a diff you approve.

Three levels of verification 🔗

For work where correctness is non-negotiable, TeXRA connects three layers of checking in one place, with agents that move freely between them.

Increasing rigor
  1. 1 LLM prose and equations Fast and creative, but unreliable on algebra.
  2. 2 Wolfram algebra and numerics Catches sign errors; verifies limits and numerics.
  3. 3 Lean 4 formal proof Machine-checked. No gaps possible.

What it has produced 🔗

Using this pipeline, a human-guided multi-agent workflow co-designed a Lean-certified catalogue of 14,116 quantum error-correcting codes with prescribed transversal diagonal gates, and resolved transversal-gate problems that had resisted manual search. The work appears in arXiv:2510.20728, and the broader case for agents in physics research is in the NeurIPS 2025 position paper arXiv:2506.06214.

Model-agnostic, and where to get it 🔗

TeXRA pairs a flagship reasoner for orchestration with cheaper, faster models for routine sub-tasks, and runs against every major provider. It ships as an extension for VS Code, Cursor, and Windsurf, and as a terminal CLI (@texra-ai/cli), published on the VS Code Marketplace, Open VSX, and npm.