An open-source AI theorist for VS Code and the terminal. It attempts real theory work: deriving results, checking derivations, formalizing proofs in Lean 4. It takes on open problems in long autonomous runs, with a team of specialist agents. Apache 2.0, and you bring your own model keys or subscriptions.
Development on main targets TeXRA 1.0, with a redesigned interface
organized around projects and a breaking change to stored application
state. Version 1.0 will start with fresh state; it will not migrate previous
JSON-based settings, session histories, or execution checkpoints. Existing
research files are outside this storage change. The implementation follows
the 1.0 development direction, using SQLite
for application state and Effect for asynchronous execution. This describes
work in progress, not the currently published release.
# VS Code (or Cursor, Windsurf, Antigravity)
code --install-extension texra-ai.texra
# Terminal: requires Node.js 22.19.0 or later in 22.x, or Node.js 24 or later
npm install -g @texra-ai/cli
# Or via Homebrew (macOS / Linux)
brew install texra-ai/tap/texraSet <PROVIDER>_API_KEY to use your own credentials, sign in with a
ChatGPT or Grok subscription, or add the API key for a Kimi Code or
GLM Coding Plan subscription.
Every agent ships bundled, including the workflow agents the
Physicist, Mathematician and Computer Scientist teams use, so every
team works offline with no sign-in. Each agent is a YAML file you can
read, copy into a custom agent and change. Signing in to a TeXRA
account (GitHub or Google, through the Profile view in VS Code or
texra login) is optional and adds no agents; model calls run on
your configured model credential.
TeXRA is sustained by the community. If TeXRA helps your research, consider supporting it via GitHub Sponsors or Buy Me a Coffee to keep it open for everyone.
In VS Code: open a .tex file, click the TeXRA icon, pick
Orchestrator or another agent, type a task. The Setup Wizard runs
on first launch and checks your environment.
In the terminal:
texra chat # interactive tool-use session
texra run polish --input paper.tex # one-shot workflow
texra multi-agent run physicist --instruction "Check this derivation" # named teamRun history and agent settings are shared between both surfaces.
Five built-in presets cover the most common research disciplines:
| Team | Built for |
|---|---|
| Physicist | Analytical derivations, numerical experiments, literature search, slide drafting |
| Mathematician | Attacking open problems, proofs, Lean 4 formalization, LaTeX correction |
| Computer Scientist | Algorithm design, code-driven experiments and ablations, tests, literature |
| Lean Project | Mathlib search, tactic simplification, blueprint-driven formalization |
| Software Engineer | An engineer lead delegating implementation, review, debugging, and testing |
Pick a team in Settings → Agents → Teams, or with texra multi-agent run <preset>. Or define your own team in YAML.
Workflow agents write to disk and produce reviewable diffs:
polish, correct, merge, ocr, transcribe_audio, paper2slide,
paper2poster.
Tool-use agents work conversationally with file, shell, and search
access: research, numerics, review, presenter, latexFixer,
latexDiff, creator, lean, assistant, setup, plus a
software-engineering line — engineer, coder, codeReviewer,
testEngineer, codeSimplifier.
Hosted specialists (signed-in users): orchestrator, search,
simplifier, criticize, firstread, logic, notation, enhance,
elevate, humanize, devise, apply, verifyFix, generic,
progressCheck, and the Lean line — leanOrchestrator,
leanBlueprint, leanSearch, leanSimplifier.
Every tool call is gated by per-stream approval. Optional YOLO mode skips approval for autonomous runs.
Bring your own keys for OpenAI, Anthropic, Google Gemini, DeepSeek, xAI Grok, Moonshot Kimi, Alibaba Qwen, Zhipu GLM, MiniMax, OpenRouter, or any OpenAI-compatible endpoint. Each agent in a team can run a different model — pair a flagship reasoner for orchestration with cheaper, faster models for routine sub-tasks.
In the extension, run TeXRA: Set API Key (stored in VS Code's
encrypted SecretStorage) or add a workspace .env:
OPENAI_API_KEY=…
ANTHROPIC_API_KEY=…
GOOGLE_API_KEY=…In the CLI, export the same variables in your shell.
- VS Code 1.125+ (also runs in Cursor, Windsurf, Antigravity), or Node.js 22.19.0 or later in 22.x, or Node.js 24 or later for the CLI
- LaTeX distribution (TeX Live, MiKTeX, or MacTeX)
- Perl (for
latexindentandlatexdiff) - Optional: ImageMagick + Ghostscript (for PDF/image processing),
git,gh, Codex CLI, Lean 4 +lake
The Setup Wizard checks for and helps install most of the above.
- Installation
- Quick Start
- Built-in Agents
- Polish a draft — workflow example
- Models
- Custom Agents
- Work produced with TeXRA
Full docs at texra.ai/guide.
Issues and feature requests: GitHub. Contact: contact@texra.ai.
Licensed under the Apache License, Version 2.0. © TeXRA Team 2024–2026.