OpenAI Codex release notes: 0.147 alpha adds auto review and MCP controls
Part of the series: OpenAI Codex release notes

OpenAI Codex alpha now has a middle path between manual approvals and letting the agent run without a review layer. The new --approve-for-me flag sends approval requests to an automatic reviewer inside a workspace-write sandbox. The same alpha line can also control where each MCP server's tools are exposed.
These are preview features, not a stable upgrade. On August 4, 2026, npm's alpha tag pointed to 0.147.0-alpha.6.4, while latest remained 0.146.0.
OpenAI Codex release notes: two new control points in 0.147 alpha
Codex CLI is OpenAI Codex running locally in a terminal, where it can read, edit, and test code. An alpha build is a preview that may change before a stable release.
The two most useful changes in this release window are:
--approve-for-meis available for interactive runs andexec. It routes approval requests through automatic review and setsapproval_policy="on-request"with theworkspace-writesandbox.- MCP configuration now accepts
omit_tools_from. A server can be hidden fromdirect,deferred, orcode_modewithout disabling it everywhere.
MCP, or Model Context Protocol, is a standard way to connect an AI agent to tools and data sources. Here, direct means a tool is present in the model's initial tool list, deferred means tool search can discover it later, and code_mode means Code Mode scripts can call it.
Sources: Codex CLI 0.147.0-alpha.6.4 on GitHub.
Source: npm registry version data for @openai/codex.
--approve-for-me reviews requests without removing the sandbox
An approval gate is the point where Codex needs permission before an action can continue. The new flag keeps that gate and routes the first review to an automatic reviewer.
--approve-for-me uses workspace-write, so Codex can write inside the workspace while the sandbox remains in place. Requests follow the on-request policy. If you already test the alpha build, you can verify the flag and then start a bounded session:
codex --help
codex --approve-for-me
This is a human startup action, not part of the agent prompt. The official CLI help describes the flag as routing approval requests through automatic review with the workspace-write sandbox.
Source: OpenAI Codex PR #36373 for --approve-for-me.
MCP tools can appear on the right surface
It used to be difficult to keep an MCP server available in one workflow but absent from another. omit_tools_from makes that distinction explicit. The following line in a server's MCP configuration hides its tools from the model's initial list and from Code Mode while leaving them available to later tool search:
omit_tools_from = ["direct", "code_mode"]
The field accepts any combination of direct, deferred, and code_mode. A large tool catalog can remain searchable without filling every session with every tool. A connector can also stay available to one deliberate workflow without being callable from arbitrary Code Mode scripts.
Source: OpenAI Codex PR #36781 for MCP tool exposure surfaces.
What Nordic teams should test now
Start with a reversible task in a real repository, such as updating tests or documentation. Keep the alpha build separate from the team's stable default, then log the version, sandbox mode, active MCP surfaces, and the requests that automatic review approved or denied.
Put integration API keys in environment variables or a secret manager. Pair that with scoped permissions and human review of the diff before merge. Codex can then do useful work without placing passwords in the prompt or turning the trial into something the team cannot audit.
Short example: use the new Codex feature
Human step: confirm that the client is 0.147.0-alpha.6.4, start the test session with codex --approve-for-me, and then paste this prompt:
Inspect this repository's current Codex and MCP configuration. Propose a minimal diff that keeps our chosen MCP server discoverable through tool search, hides it from direct and code_mode, and lists which approvals must remain human-owned. Change nothing.
Good output should:
- Name the real configuration file and affected MCP server.
- Place
omit_tools_from = ["direct", "code_mode"]correctly. - Separate automatic review from final human approval.
- Leave files and settings unchanged.
If this test becomes a recurring routine, Hammer's Tool Forge can package the configuration, permissions, and run log so the team does not rebuild the same workflow for every user.
What to watch next
0.147.0-alpha.6.4 is still an alpha build, and npm's stable latest tag remained on 0.146.0 when checked. Treat these features as a documented test rather than a silent change to the team's default. The next stable Codex CLI release will show which names and behaviors carry forward.
FAQ
Is Codex CLI 0.147.0-alpha.6.4 a stable release?
No. It was on npm's alpha tag on August 4, 2026, while the stable latest tag still pointed to 0.146.0.
What does --approve-for-me do in OpenAI Codex?
It routes approval requests through automatic review and uses the on-request policy with the workspace-write sandbox. It does not remove the sandbox.
Which MCP values does omit_tools_from accept?
The field accepts direct, deferred, and code_mode in any combination for the relevant MCP server.
The Forge newsletter
Get new articles in your inbox
Pick the topics you care about. No noise, at most one email a week.
We follow GDPR. Unsubscribe anytime.


