AI

Configure Codex chats

Choose access, models, reasoning, speed, defaults, and parallel workers.

Start with Open Codex chat if you have not used the agent before.

Understand Codex access

New chats created through the hosted CoCalc UI use full project access: Codex can read and modify files, run commands, install software, and use the network inside the project environment. Explicitly read-only sessions can still be created through other supported interfaces. A working directory selects where commands start; it is not a permission boundary.

  1. Open the project whose files and software Codex should work with.
  2. Open Codex settings using the Codex control. On phones, tap the button showing the model and reasoning level in the chat header.
  3. Read the Access notice and include the files, intended changes, and constraints in your request.

CoCalc Lite offers an Access choice instead: Read only, Workspace write, or Full access. These configure the Codex command sandbox. Read only still permits network access. Workspace write permits workspace and temporary-directory writes; Full access removes that command-sandbox file restriction. Select the mode appropriate for the task and click Save.

Choose model, reasoning, and speed

Open Codex settings using the Codex control. On phones, tap the button showing the model and reasoning level in the chat header. These settings apply to the selected thread; the compact chat controls show the same model and reasoning level.

  1. Under Model and session, select a Model.
  2. Choose a Reasoning level from the options supported by that model. Recheck it after changing models, because the available levels and default can change.
  3. Choose Standard or Fast under Speed, then click Save.

Fast is available only for supported models and uses more Codex credits. Choose it when lower latency is worth the higher usage; Standard is the default. Changing to a model without Fast support returns speed to Standard.

Membership-funded turns use the model, reasoning, and speed selected by CoCalc, so those controls are constrained. Connect and select a personal ChatGPT plan or OpenAI API key when you need other available model settings. These choices configure future turns rather than rewriting earlier replies.

Set defaults for new chats

To reuse your preferred model and reasoning level, open account AI settings and find New Codex chat defaults.

  1. Choose Model and Reasoning. The reasoning choices depend on the selected model.
  2. In Lite, also choose the default Execution mode. Hosted CoCalc shows its full-project-access notice instead.
  3. Click Save defaults.

The saved defaults are used when you create a new Codex chat. They do not serve as a bulk edit of existing threads: open a thread's Codex settings to adjust that thread. Use Reset to built-in defaults when you want new chats to start with CoCalc's supplied choices again.

Review the selected thread's settings before starting a task, especially if you changed its payment source. Membership-funded turns use the settings selected by CoCalc even when your account defaults request another model.

Set the parallel subagent preference

Codex can use worker agents alongside the manager handling your request. The Parallel subagents button in Codex settings opens the same preference as Maximum concurrent subagents in account AI settings.

  1. Open either control.
  2. Choose Automatic (currently 3) or a number from 1 through 16.
  3. Allow the setting to take effect when the Codex session is next loaded.

The number describes worker agents in addition to the manager. It is an account preference used to configure sessions, not a request to immediately create that many workers, and not a displayed total for every session on the account.

Higher values can consume your Codex or API allowance much faster. Start with a value appropriate for the work and inspect session activity if you are unsure what remains running. Changing the preference does not by itself stop existing workers; use the session's stop controls when you need to interrupt ongoing work.

Refresh the models available to your ChatGPT account

When a thread uses your ChatGPT Plan, CoCalc can check which models that account currently supports.

  1. Open Codex settings using the Codex control. On phones, tap the button showing the model and reasoning level in the chat header.
  2. Under Model and session, click Refresh models and wait for the check to finish.
  3. If Model unavailable for this ChatGPT account appears, choose an enabled model. CoCalc keeps the previous selection visible instead of silently replacing it.
  4. Recheck Reasoning level and Speed, then click Save before starting the next turn.

Refreshing checks availability; it does not grant access to additional models. A model list can contain cached or built-in choices when a live check has not succeeded.

Refresh models is specific to ChatGPT-funded threads. API-key and membership funding do not use this account-model refresh control.