Skip to content

Tailoring TeXRA ​

The assistant's own prompt and tools are fixed. You shape what it does in a project with the things below; none of them replaces the assistant, they add to it.

Project instructions: AGENTS.md ​

Put the rules every conversation in a project should follow (notation, macros, journal style, how to build the paper, files never to touch) in an AGENTS.md file at the workspace root. TeXRA adds it to the assistant's instructions in every conversation. Read Project instructions.

Memory ​

The assistant can keep notes about the project in its memory and read them in later conversations. Read Memory.

Plugins on and off ​

Settings → Plugins (or /plugins in the terminal chat) switches each group of tools on or off: web search, Zotero, Lean, Wolfram, the Codex and Claude Code integrations, multi-agent work and more. A plugin that is off adds nothing to the conversation.

Skills ​

A skill is a folder with a SKILL.md that tells the assistant how to do one kind of work. TeXRA reads skills from installed plugins, from .texra/skills/ in the project, from your own skills folder, and, once you switch that source on, from the .agents, .claude, .codex and .gemini skill folders beside the project and in your home folder. Turn skills on with Enable skills for agents in the Skills section of Settings → Agents, where you also choose the sources. In the terminal, texra skills list shows them and /skills applies one to your next request. Read Skills in the CLI guide.

Install a plugin ​

TeXRA installs plugins published for Claude Code and Codex. A plugin can bring skills, commands, agents, MCP servers, programs (programs/<name>.js), and native TeXRA extensions (harness.mjs):

bash
texra plugin install github.com/<owner>/<repo>
texra plugin enable <name>

An installed plugin stays off until you review and trust what it carries (a native harness.mjs entry loads from a snapshot of the trusted files, and any edit requires re-trusting). In the terminal chat you can also use /plugins add, and in the VS Code extension and desktop app manage installed plugins under Settings → Plugins. Read Plugins in the CLI guide.

A plugin's agents are Claude Code subagent files (agents/<name>.md: a name, a description and an optional tool list, with the agent's instructions as the body). While the plugin is enabled, the assistant is told of each one as <plugin>:<name> and can hand it a piece of work by that name, so ask for it: "have my-plugin:referee read section 3". An agent that lists no tools gets the file, shell and web tools, the tools every agent gets injected (plan, script, context, and the rest), and the tools of the plugins you have on.

Write your own program ​

A program is a JavaScript file that runs a fixed procedure over the files you give it. Put it in a plugin's programs/ folder and it is started as <plugin>:<name>, by you (texra run <plugin>:<name> --input …) or by the assistant when you ask for it. Read Programs in a plugin.

The assistant can also write a program for itself, saved in the project's memory, and start it again in later conversations. Read Programs an agent writes for itself.

MCP servers ​

The Model Context Protocol servers an installed plugin carries give the assistant their tools while the plugin is enabled. Read MCP servers.

A different point of view is a sentence ​

You need no new agent to change how a job is done. The programs take their scope and point of view from the instruction: "read as a referee for a physics journal and mark every unsupported claim" is annotate with that sentence. To combine steps, ask the assistant to chain programs, for example annotate, then apply.

If you used 0.40

Custom agent YAML files, the custom agents folder, the Agents library in Settings, customizing a built-in agent, and TeXRA's bundled skills are gone. Move an agent's standing rules into AGENTS.md, a checklist into a skill, and a repeatable procedure into a plugin program. The skill collections listed on the Open source page are published as Claude Code plugins.