Skip to content

Open-source multi-agent AI for theorists

Derive it, check it,
do good work.

You tell one assistant what you need. It searches real databases, computes in Wolfram, proves in Lean 4, splits large jobs across helper agents and programs, and hands back every change as a diff you approve. Open source under Apache 2.0, in VS Code, its forks, and the terminal.

or run it anywhere from your terminal
$ npm install -g @texra-ai/cli
spectral-gap — texra-paper

2  Preliminaries

Definition 1.An Erdös-RényiErdős–Rényi random graph G(n, p) is a graph on n vertices where each posible edgepossible edge {u, v} is included with probability pindependent, independently of all other edges.

Let A be the adjacency matrix of a graph G be a graph on n vertices. Its adjacency matrix A = A(G) is an n × n matrix where Auv = 1 if {u, v} is an edge in G. The eigenvalues of, and Auv = 0 otherwise. Since Aare denotedis symmetric, its eigenvalues are real and are denoted by λ₁ ≥ λ₂ ≥ ⋯ ≥ λₙ.

3  Spectral Gap in Random Regular Graphs

For a d-regular graph G (where every vertex has degree d), it is well-known that the largest eigenvalue of its adjacency matrix is λ₁ = d. The spectral gap, defined as d − λ₂, plays a crucial role in the expansion properties of the graph.

Theorem 2 (Alon–Boppana).For a d-regular graph on n vertices, λ₂ is largeλ₂ ≥ 2√(d−1) − o(1) as n → ∞. A d-regular graph meeting this bound with equality is called a good expanderRamanujan graph.

Proof. It follows from counting walks.Count closed walks of length 2k rooted at a fixed vertex. The number of such walks in the d-regular tree is the Catalan-weighted moment Ck (d−1)k, and comparing with tr(A2k) = Σi λi2kas n → ∞ forces λ₂ ≥ 2√(d−1) − o(1).∎

Output you can check

Derivations are checked in Wolfram, proofs compile in Lean 4, citations resolve to real database entries. Edits arrive as diffs you read line by line before they touch your files.

One assistant, many hands

You tell one assistant what you need. It splits a large job across helper agents and programs that search, compute, review, and formalize, each with only the tools its piece needs.

Your keys, direct to the model

Use your own API keys or existing ChatGPT, Grok, Kimi Code, or GLM subscriptions. Requests go straight from your machine to the provider.

One assistant, programs for every stage

You describe the problem. The assistant breaks it into sub-tasks, runs helper agents and programs in parallel, and returns results you approve before they touch your files.

Explore
search
arXiv, Zotero
Derive
derive · experiment
Wolfram, bash
Verify
review · prove
Wolfram checks, Lean proofs
Write
edit · annotate · apply
LaTeX, figures, diffs
Present
slides · poster
Beamer, poster
Document tasks
Read project→Plan→Draft
Reflect→Output
.tex in → improved .tex + latexdiff
Assistant
Breaks work into sub-tasks, starts helper agents and programs, coordinates results
Chat
Your question→Tool calls→Reason
Answer
Conversational · persistent across sessions
leanarxivbashwebzoteroreadwritegrep25+ built-in tools

How researchers use TeXRA

Derive

Steady-state entanglement in a driven spin chain

The assistant (or the derive program) builds the Lindblad superoperator, solves for the steady state in Wolfram, computes the concurrence analytically, and cross-checks with exact diagonalization in Julia at N=8.

Write

Finalizing a 40-page paper on spectral graph theory

The edit program unifies notation (\lambda_2 vs \mu for the same eigenvalue across sections), fixes label conflicts, and outputs a diff you review line by line.

Verify

Formalizing subadditivity of entropy in Lean 4

The assistant (or the prove program) searches Loogle for the right Mathlib lemma, reads the proof state, adds a missing hypothesis, and produces a proof that compiles with zero errors.

Start in under two minutes

Install the VS Code extension from the Marketplace or the CLI (npm install -g @texra-ai/cli); add your own API key or connect a provider subscription, and run your first task.

terminal
$npm install -g @texra-ai/cli
$texra auth chatgpt login
$texra run edit --input paper.tex
executions/9f3a6c81d24e/r1/paper.tex

Common questions

What models does it support?

OpenAI, Anthropic Claude, Google Gemini, DeepSeek, xAI Grok, Moonshot Kimi, Qwen, GLM, MiniMax, Meta Muse Spark, and any OpenAI-compatible endpoint. Bring your own API key, or use a provider subscription you already pay for (ChatGPT, Grok, Kimi Code and the GLM Coding Plan). Each helper agent or program the assistant starts can run a different model.

Does it work with Overleaf?

Yes, via git sync. Read the Overleaf guide.

Does it work with Lean 4?

Yes: Loogle search, proof state inspection, diagnostics, build and cache management. TeXRA spawns its own Lean server in every build (VS Code, desktop and CLI), so you only need elan (lake on PATH) and a Lake project.

Can I use it without VS Code?

Yes. The @texra-ai/cli terminal client runs the same assistant, programs, and sign-in on your .tex projects for scripts, CI, and remote machines. Read the CLI guide.

Is my data private?

Yes. Model calls go directly from your machine to the provider, whether you use your own API key or a provider subscription. TeXRA does not sit between you and the model. TeXRA sends anonymous usage metadata (built-in program name, model, token counts), never prompts or document content, and you can turn it off; read Usage logging.

Is TeXRA open source?

Yes. The extension, the CLI, and the agent runtime are published under the Apache 2.0 license, and you run them with your own model keys or subscriptions. Read the open source guide.

Can I tailor TeXRA to my project?

Yes. Put your project's conventions in AGENTS.md, add skills, and install plugins that carry skills and programs of your own. Read Tailoring TeXRA.

Where do I get help?

Email contact@texra.ai or open an issue on GitHub.