Use Claude Code and Codex together

Connect both coding providers

Install the official Claude Code and Codex CLIs, then open More, choose Providers, and connect each account from Accounts. KalCode opens the provider's own sign-in flow and starts each agent with the account, models and effort choices that provider actually offers. Your provider subscriptions, organization policies, quotas and native approvals still apply. The providers reference explains what is preserved from a normal terminal session.

Open the project in Code. New agent starts the project's last valid setup immediately; use its options to choose Claude Code for one launch and Codex for the other. Each opens as a separate interactive terminal. Neither provider is automatically the lead: assign roles for the task based on the change you need, the accounts available and the workflow you want to supervise.

Use a bounded implement-and-review loop

The safest first workflow has one writing agent and one reviewing agent. Let the implementer finish a coherent candidate before the reviewer starts inspecting it. Ordinary New agent panes share the opened project folder, so asking both to edit the same files at the same time creates avoidable races. Use isolated worktrees when you need simultaneous implementation.

1. Give the implementer one outcome, an owned file area and exact validation.
2. Wait for a reviewable diff and the implementer's test result.
3. Tell the second agent to review without editing.
4. Return confirmed findings to the implementer, then rerun only the affected checks.
5. Inspect the final diff and decide when to commit, merge or deploy.

Prompt the implementing agent

Use a task with a boundary and a finish condition, for example:

Own the password-reset form and its focused tests.
Reproduce the double-submit bug, identify the root cause, and make the smallest fix.
Preserve the existing API contract and accessibility behavior.
Run the form tests and report the files changed, commands run, and any remaining risk.
Do not merge or deploy.

This prompt works with either Claude Code or Codex. It does not assume one model is better, and it leaves repository-wide decisions outside the agent's scope.

Prompt the reviewing agent

Start the review after the candidate exists. Ask for concrete, reproducible findings rather than a general opinion:

Review the current password-reset diff. Do not edit files.
Check correctness, regressions, security, accessibility, and whether the tests prove the fix.
For each real issue, give severity, file and line, reproduction, and the smallest correction.
Say explicitly if you found no blocking issue. Do not merge or deploy.

The reviewer may run read-only inspection commands and relevant tests through its native tools. If it reports a problem, reproduce it before changing code. A second agent's approval is useful evidence, but it is not proof that the branch is ready.

Keep the handoff inspectable

Put the accepted finding back into the implementing terminal with the same boundary: fix that issue, add or adjust focused coverage, and report the new result. Compare the final diff with the original request. Agent state such as Done only describes the provider session; it does not mean verified, merged or shipped.

Native provider permission prompts still appear in the provider terminal, and access to credentials or secrets follows KalCode's permission rules. Do not paste credentials into a prompt to get around either system.

Scale beyond two agents carefully

When tasks can be split by files or components, move to the parallel coding agents workflow. Squads can declare owned paths, dependencies and isolated Git worktrees; simple New agent panes remain useful for sequential implementation, investigation and review in the same project folder. If you have not installed KalCode yet, use the official download page, then follow the setup guide to connect both providers.