Codex

Drives replies through the Codex CLI installed on your machine, using its ChatGPT login instead of an API key. Frontier OpenAI models at a flat subscription price.

Best suited to the Assistant role rather than to chat: startup costs a moment on every reply. It can also serve text generation, summarization, action inference and computer vision.

Experimental and admin-only. It drives your own OpenAI account; automated or heavy use outside OpenAI's intended workflows may break their Usage Policies and get the account suspended. Explicit content is not supported.

Setup

Install and log in

Install the Codex CLI and run codex login once in a terminal, so it works on its own.

Add the service

Manage Modules → Add Modules → Codex → Install. Point Codex Executable at the CLI if codex is not on your PATH.

Pick a model

The model list comes from the CLI, so it shows what your plan offers. Leave it empty to use the CLI's default. Then set Codex as your Assistant service, and/or your text generation service.

Settings

Setting	Notes
Model	From the CLI. Empty uses the CLI's default. Picking one sets the max context to 272K tokens.
Reasoning Effort	Levels reported by the CLI. Default Medium; the vision preset uses Low. Turning Enable Reasoning off asks for the lowest effort, since Codex models always think.
Let the model see images	On: images go to the model. Off: your Computer Vision service describes them first.
Keep the thread between turns	On by default. Each reply continues the previous Codex thread and sends only what is new. Applies to Assistant chats and Multi-Turn presets. An edit or a summary starts a new thread; a regeneration rolls it back one turn.

Agentic tool use

Off by default. Turning it on lets an Assistant chat run Codex with its real tools — read and edit files, run commands — under the Agentic Working Directory (default Data/Agents/OpenAICodex/Working). Other chat styles and the scenario assistant keep Voxta's own tools only, so a character cannot reach your disk mid-scene; summarization and action inference stay text-only.

Tool Permissions: Ask (default, asks before anything beyond reading), Plan (read-only, ends with a plan to approve), Allow Edits, or Allow All. Outside a chat, anything above the level is declined.
Review Risky Tool Calls (on, see Tool permissions & review): before a destructive call runs without asking, your action inference service reviews it. Obviously destructive commands, writes to system folders and access to saved credentials are always refused.
Use Codex's native harness: keeps Codex's own agent instructions and the working directory's AGENTS.md, with the character prompt layered on top.
The chat's tools are passed to Codex over MCP.

This lets a language model read and write files and run commands on your machine. Leave it off unless you specifically want that, and set the working directory deliberately.