OpenAI Codex release notes: the ChatGPT app and writes mode in 0.144

OpenAI Codex is moving closer to the everyday work surface. On July 9, Codex joined the ChatGPT desktop app for macOS and Windows, while Codex CLI 0.144.0 added a new writes approval mode. Patch 0.144.1 then fixed installation and code-mode reliability. The pieces fit together: more work can happen in one project view, while write actions can still pause for a human decision.

OpenAI Codex release notes: write boundaries meet a broader app

Codex CLI is OpenAI's coding agent that runs locally and can read, change, and test code in a project. Version 0.144.0 added an app approval mode called writes. Tools explicitly marked as read-only can run without interruption. Every other tool prompts before continuing, including non-destructive writes and tools without the required annotation.

The same release made interactive authentication for MCP tools stable and enabled by default. MCP, or Model Context Protocol, lets an AI client connect to external tools and data sources. Users and clients no longer need an experimental opt-in before an MCP tool can request authentication.

Source: OpenAI Codex changelog, CLI 0.144.0

Source: GitHub PR #30482 for writes mode

Codex is now in the ChatGPT desktop app

The second change is the work surface. Codex is now part of the ChatGPT desktop app on macOS and Windows. Existing Codex app users can update while keeping their projects, settings, and workflows. The app supports direct Markdown and code editing, inline annotations, and revisions to selected content. GitHub pull requests can be reviewed in the sidebar, and one project can span multiple repositories.

That helps when one change touches a website, an integration, and internal documentation. Keeping several repositories in one project cuts down on separate sessions. It also makes the boundary around write actions more useful: inspection can continue, while changes can still require approval.

Source: OpenAI Codex changelog, Codex joins the ChatGPT desktop app

0.144.1 is the current stable fix

Codex CLI 0.144.1 followed 0.144.0 with a focused patch. It fixes standalone installs when GitHub returns compact or reordered release metadata. On macOS, package installs now expose the code-mode host beside codex, and code mode falls back to the embedded runtime when the companion host binary is unavailable.

This patch does not introduce another operating model, but it is the version to use when testing the changes above. The npm registry lists 0.144.1 as the current latest release.

Source: OpenAI Codex changelog, CLI 0.144.1

Source: npm registry for @openai/codex 0.144.1

Human step: update the CLI before testing

OpenAI's release notes provide this installation command for the current stable version:

npm install -g @openai/[email protected]

Keep package installation and configuration changes as human actions. The prompt below should not update packages, change approval mode, or connect new tools.

Source: OpenAI Codex changelog, 0.144.1 installation line

Short example: use the new Codex feature

Inspect this multi-repository project without editing files. Map one requested change across the repositories, list the read-only checks you completed, then name every write action that should require human approval. End with a short review plan and do not execute the writes.

Good output should:

• clearly separate completed reads from proposed writes
• show which repositories are actually affected
• identify approval points before changing files, configuration, or external systems
• finish with a short review order a person can follow

What a Nordic team can test now

Choose a workflow that already touches two repositories, such as a website and its integration layer. Let Codex map the change without writing. Then verify that your tool definitions correctly annotate read-only actions before rolling out writes mode more broadly. A missing or incorrect annotation should cause a prompt rather than a silent assumption.

For live integrations, keep authentication in environment variables or a secret manager, use scoped permissions, and log approved writes. In Tool Forge, we can connect the tools, define where write actions must stop, and make each change reviewable.