AI Agent Settings

Control how the assistant responds, which tools it can use, and where it runs.

Open the settings from the three-dot menu in the assistant panel and choose AI Setup. They live in the AI group of the settings dialog. Some tabs apply to your account and some to the current project.

The three-dot menu of the assistant panel with the AI Setup entry highlighted
AI Setup is the first entry in the assistant's three-dot menu.

Account tabs

Style

These settings shape how the assistant writes back.

  • Planning: Planning is the default. The assistant writes a plan first and follows it for every task. Off is one-shot: no plan, it acts straight away. Use Planning for multi-step requests and Off for quick back-and-forth.
  • Verbosity: Brief, Medium or Verbose.
  • Tone: Casual, Balanced or Formal.
  • Language level: from Simple to High-level. High-level means scientific and technical vocabulary with no simplification.
  • Autonomy: Educational means the assistant teaches and handles mundane tasks only, while you do the writing and decisions. Agentic means it drives the work end to end with minimal check-ins.
The Style tab of AI Setup with Planning, Verbosity, Tone, Language level and Autonomy selectors
The Style tab.

Rigour

Rigour mode (experimental) has an independent AI review every edit and the final reply against the real session evidence before it reaches you. It raises quality at the cost of slower responses. You choose which checks apply:

  • Traceable sources
  • Full-text verification
  • No lingering doubt
  • Limitations disclosed
  • No hidden assumptions
  • Proportionate answer
  • Won't need revisiting
The Rigour tab with the Rigour mode toggle and the list of checks
The Rigour tab: the mode toggle and the checks you can choose.

Tools

Core tools are always available. On-demand tools, such as web search, citation and arXiv search, sub-agents and cross-project access, are loaded only when a task needs them. Turn one off to stop it being used at all this session.

Local AI Server

Routes chat requests to your own Ollama instance. Turn on Enable Local AI Relay, set the Ollama Host URL, and choose a model for each of Fast, General, Smart and Summary. See Using Ollama.

The Local AI Server tab with the relay toggle, the Ollama host URL and a model field for each tier
The Local AI Server tab.

BYOK

Use my own OpenRouter key replaces the Smart model with your own OpenRouter key and model, for example anthropic/claude-sonnet-4.5. OpenRouter bills you directly, and these requests do not count against your TeXposit credit limit. The key is never shown again after saving. The tab also shows tokens used and an estimated cost based on OpenRouter's published prices.

The BYOK tab with the OpenRouter key toggle, API key and model fields and the usage counters
The BYOK tab.

Personal instructions

Free text that only you see. It applies in every project you open, for example "Explain LaTeX errors in plain terms before fixing them."

The Personal instructions tab with a text box and the Apply instructions in this session checkbox
The Personal instructions tab.

Project tabs

General

Enable AI assistant for this project turns off the chat assistant, inline suggestions, selection edits and auto-fix for the project. Skills and instructions stay saved.

Instructions

Project instructions are shared with everyone on the project and versioned with it, stored in .texposit/instructions.md. Put conventions and boundaries here:

# Project instructions

## Conventions
- Cross-references: use \cref (cleveref is loaded).
- Citations: \citep / \citet (natbib, author-year).
- British spelling.

## Boundaries
- Do not edit preamble.tex. Ask first.

Draft from project has the assistant read your project and propose a first version. Apply instructions in this session lets you turn instructions off for one session. With viewer or reviewer access, instructions are read-only.

Skills

Skills are playbooks the assistant loads when a task matches, such as a citation style, LaTeX conventions or a writing workflow specific to the project. Create one with New skill. The name uses lowercase letters, numbers and hyphens, and cannot be changed afterwards. Deleting a skill can be undone from the project history.

Extended reasoning

Type /thinking in the chat to toggle extended reasoning, which lets the model think longer before responding. It has an effect only on reasoning-capable models.

Organisation limits

If you use TeXposit through an organisation, an administrator can disable the assistant, or the customization settings (instructions and skills), for everyone. Those controls then do not appear.