Codex MCP tool adds separate review loop for uncommitted changes

A new MCP tool for Codex runs app-server reviews of uncommitted changes in a separate context from the main coding session. The project says this reduces regressions, keeps review context clean, and supports longer autonomous runs.

Codex MCP tool adds separate review loop for uncommitted changes

A new MCP tool lets Codex review uncommitted changes in a separate loop instead of using the built-in /review command. The tool, called review_uncommitted_changes, is designed to keep review context isolated while the main session continues making changes with the implementation details still available.

The project behind the tool says that split helps reduce regressions and improves code quality during longer autonomous runs. It also reviews uncommitted changes, including untracked files, using Codex’s native app-server review target uncommittedChanges.

⚡ New to this?

This is news about a tool that changes how an AI coding assistant checks its own work. MCP stands for Model Context Protocol, a way for tools and assistants to talk to each other through a standard interface. The practical change here is that code review can happen in a separate session, which can make the assistant less likely to lose track of the original implementation details.

🦞 OpenClaw angle

If you run self-hosted coding agents, separate “write” and “review” into different tool paths instead of asking one session to do everything at once. Keep review runs read-only, set timeouts higher than the agent’s own review window, and make the review tool consume uncommitted changes so it sees the real working tree, not just committed code. If you use MCP, register the review server through your normal config so you can swap models or prompts without changing the agent logic.

The setup requires the Codex CLI to be installed and authenticated, along with uv. Codex loads MCP servers from ~/.codex/config.toml, and the supported way to add the server from the command line is codex mcp add, according to the project instructions.

One example configuration uses a review profile with gpt-5.5 and xhigh reasoning effort. The project also recommends setting tool_timeout_sec in ~/.codex/config.toml to a value higher than the review timeout passed to the server command.

In the example, if the review timeout is 2,700 seconds, tool_timeout_sec is set to 3,000. The server is registered as a stdio MCP server that Codex launches when a session starts.

The repository provides several ways to run it. Users can add the tool from a local clone with uv run -m mcp_code_review.server, pull it directly from GitHub with uvx --from git+https://github.com/Szpadel/codex-mcp-code-review, or point uv at a local project path.

The default configuration uses four runs, though that can be changed with the MCP server’s --parallelism setting. The server also runs in a read-only sandbox with approval policy set to never, according to the source.

Additional review instructions are enabled by default. Existing setups that already use --enable additional_review_instructions still work, but users can turn the feature off with --disable additional_review_instructions.

When the extra instructions are enabled, the server changes the review run to a custom prompt based on a synced copy of Codex’s native uncommitted-changes instructions and adds an “Additional review instructions” section. If the feature is disabled, the tool schema does not expose the additional_developer_instructions argument, and requests that send it anyway are rejected.

The tool is available through the codex-code-review MCP server entry, and the project says the CLI workflow is the supported way to add MCP servers. The repository for the tool is published at https://github.com/Szpadel/codex-mcp-code-review.

Source: HN Show HN ↗

More from OpenClaw News