TEXOR

TEXOR opens a local browser workbench from VSCode and uses TEXOR Agent, Codex CLI, or Claude Code as the actual paper-writing agent.

The workbench is designed for manuscript iteration:

• Load an existing research project or create a new TEXOR project.
• Import an existing .tex manuscript as the first paper version.
• Compile and compare previous/current PDF versions.
• Mark text or regions in the PDF, write a natural-language revision note, and let the chosen agent update the LaTeX.
• Keep agent progress visible while TEXOR stores each completed revision as a new version.

Command

Run one command from the VSCode Command Palette:

TEXOR: Open Browser Workbench

The extension starts the local TEXOR web and API services, then prompts you to open the workbench in your browser.

Requirements

• TEXOR Agent can use any OpenAI-compatible API through baseUrl + apiKey + model.
• Codex CLI is optional and can still be used through texor.codexExecutable or the OpenAI ChatGPT/Codex VSCode extension.
• Claude Code is optional and can still be used through texor.claudeExecutable.
• A LaTeX engine such as pdflatex or lualatex is required for PDF compilation.

Agent Backends

TEXOR supports three agent backends:

• TEXOR Agent: TEXOR runs its own lightweight state-graph loop, calls the user's OpenAI-compatible model API, and executes explicit tools such as reading project files, writing main.tex, listing files, and running non-destructive project commands.
• Codex CLI: TEXOR delegates the task to a local Codex CLI session.
• Claude Code: TEXOR delegates the task to a local Claude Code session.

The built-in TEXOR Agent is intentionally modeled after LangGraph-style state machines: model step, tool step, state update, checkpoint through the project command/session, then final save. It avoids bundling LangGraph in the extension for now so the VSIX stays small and easy to debug.

Task Routing

TEXOR Agent now routes work by intent:

• Fast local wording polish for selected PDF text.
• Full-manuscript consistency revision when the request affects structure, claims, terminology, or multiple sections.
• Structure-diagram generation through the configured image model, saved into the project and referenced from main.tex.
• Result-figure updates by inspecting and running project-local plotting/code files.
• Reference and related-work support through online paper search metadata and abstracts.

Project And Manuscript Loading

• 加载项目 opens an existing TEXOR project record, including its Agent conversation and paper version history.
• 新建项目 creates a new TEXOR project record for a code project path.
• 导入当前项目 .tex imports an existing manuscript into the current project. TEXOR copies it to <project>/.texor/manuscript/main.tex; later revisions use that file as the manuscript entrypoint.

The code project path is required. The .tex path is optional and only seeds the current project manuscript.

Templates

TEXOR ships only a small template catalog. Full journal/conference template archives are downloaded on first use into VSCode extension storage. Supported direct downloads currently include IEEEtran, ACM acmart, Elsevier elsarticle, Elsevier CAS, and ICLR 2026.

Settings

• texor.serverUrl: TEXOR backend URL. Defaults to http://127.0.0.1:4174.
• texor.webUrl: TEXOR browser workbench URL. Defaults to http://127.0.0.1:4174.
• texor.appPath: optional override for the bundled TEXOR application directory.
• texor.targetJournal: optional default target journal or conference.
• texor.codexExecutable: Codex CLI executable path.
• texor.claudeExecutable: Claude Code CLI executable path.
• texor.claudeModel: optional Claude model override.
• texor.agentProvider: provider label for TEXOR Agent.
• texor.agentBaseUrl: OpenAI-compatible base URL for TEXOR Agent.
• texor.agentModel: model name for TEXOR Agent.
• texor.agentImageModel: image model used for structure diagrams, for example gpt-image-1.
• texor.agentApiKey: API key for TEXOR Agent.