OpenAI Codex release notes: 0.142.2 makes MCP tools easier to discover

OpenAI Codex has a small but useful Codex changelog signal after the previous post: 0.142.2 changes how Codex discovers MCP tools, how macOS clients can follow system proxy settings, and how plugins appear in dark mode. 0.142.4 is now the latest stable npm version, but the GitHub release notes say 0.142.3 and 0.142.4 have no identified user-facing changes. So the practical news is still 0.142.2.

Codex CLI is OpenAI Codex as a local coding agent in the terminal: it can read, change, and run code in the workspace you give it. MCP, Model Context Protocol, is a way to connect an AI agent to external tools and data sources with clearer boundaries. An approval gate is a point where a person or policy must approve a risky action before the agent continues.

Source: OpenAI Codex GitHub release 0.142.2 and the npm registry for @openai/codex.

OpenAI Codex release notes: what changed in 0.142.2?

The most useful change for practical teams is that MCP tools now use tool search by default when supported. That sounds dry. In daily work, it means Codex can discover the right tools in an MCP connection more reliably, while the release explicitly preserves compatibility with older models and providers.

The release notes also include several smaller operational changes:

• macOS authentication can honor system proxy, PAC, and WPAD settings when respect_system_proxy is enabled.
• Plugins can provide dedicated dark-mode logos through local manifests and remote catalogs.
• Apps can show richer safety-buffering UI with server-provided visibility and faster-model metadata.
• Remote stdio MCP servers now accept absolute working directories written in the remote platform's path format.
• PowerShell commands containing AST regions the safety classifier cannot inspect now require approval.

Source: OpenAI Codex GitHub release 0.142.2.

Why MCP tool search matters for Swedish teams

Many Codex pilots start as terminal help inside one repo. The next step is usually integration: GitHub, issue tracking, internal documents, build logs, or private APIs. At that point the agent being smart is not enough. It has to find the right tools, respect the boundaries, and leave a trail people can review.

Tool search in MCP is one of those small infrastructure changes that can make larger workflows less brittle. For Hammer readers, it matters most if you already test Codex with your own MCP servers, remote environments, or plugin catalogs. Put the effort into a simple run receipt: which tools were found, which actions needed approval, which keys or environment variables were involved, and what should be redacted from logs before sharing.

Short example: use the new Codex feature

Human step: open the workspace where you already run Codex and use your normal package or client management to confirm that you are on a build that includes the 0.142.2 changes. The npm registry lists 0.142.4 as the latest stable version, while the GitHub release notes say the later 0.142.3 and 0.142.4 releases are maintenance-only with no user-facing changes.

Source: npm registry for @openai/codex, release 0.142.3, and release 0.142.4.

Then paste a short agent instruction:

Inspect this Codex workspace for MCP and plugin readiness. List which MCP tools are discoverable, which actions would need approval before execution, and whether proxy or credential notes are documented for macOS or Windows users. Do not edit files. Return a short run receipt with risks, missing docs, and one safe next test.

Good output should show:

• which MCP tools Codex can actually see, not only what the docs mention
• which actions should pass through an approval gate before execution
• whether proxy settings, API keys, environment variables, and log redaction are documented well enough
• one next test small enough for a human to review

What to test now

If you use Codex without MCP, it is enough to note the version and move on. If you connect Codex to internal tools, 0.142.2 is a good reason to do a quick control pass: run a read-only task, ask Codex to describe which tools it can see, and require human approval before it changes files, runs PowerShell, or touches sensitive connections.

Hammer Automation would place this in Tool Forge/Verktygssmide: not as a big platform project, but as a way to make a working Codex pilot more operable. Start with tool discovery, scoped permissions, secrets through environment variables or a secret manager, log redaction, and a simple audit log for each run.