Applied AI Notes · November 2026
nouez: Claude Code Asks Codex for a Second Opinion
An open-source MCP server that lets Claude Code consult a read-only Codex session, and open Claude Code sessions on my computer from my phone.
When I need a second reader
Some steps in my work are hard to undo: a production release, deleting files, a paid experiment, publishing a note, or treating an analysis as a conclusion. Before each of them I want a second reading from a model trained by another company, which can catch what the model that did the work missed. When Codex reviewed my note on agent harnesses, it found a diagram saying the agent loop ends when the model stops asking for tools, although some harnesses keep it going, and a sentence calling fixed workflows what “most business requests” describe, when my own survey put them at 41%. Neither error would have broken anything, but both would have misled a reader.
Without a bridge, getting that reading means two terminals and copying questions and answers between them by hand. nouez removes that step. It also lets me ask a Remote Control session on my phone to open or resume another project on my computer.
How I use it
Claude Code does the work and owns it. When a step needs review, Claude starts a Codex “consultant” in the same project, sends it the relevant files and a narrow question, and keeps working while the consultant reads. The consultant’s local file access is read-only, and Claude applies any edit it proposes, so one author stays responsible for the result. I can open the consultant in a terminal pane beside Claude to watch its work. The reply wakes Claude when it arrives, and I treat the consultant’s findings as claims to check against their sources, in the same way as the draft’s.
My consultations fall into these groups, roughly largest first:
- following up on earlier consultations, often to re-check fixes;
- validating data, analyses and decisions;
- critiquing plans and launch gates;
- reviewing text and checking its claims;
- researching against outside sources;
- reviewing code or diagnosing bugs.
Many of them ask for a verdict in a fixed format, such as go, changes or no-go, followed by what the consultant did not check, so that silence is not mistaken for approval.
From my phone, I ask a Remote Control session to open a new Claude Code session for a project, either a new folder or an existing one, and it appears in the app’s session list a few seconds later. An earlier conversation in that project can be picked up again, and a session that is stuck can be listed and closed from the phone too.
What nouez is, and why MCP
nouez is a small MCP server written in Python using only the standard library, so it needs no packages. Claude Code talks to it over the Model Context Protocol, and it talks to the Codex app-server daemon on the same computer, where the consultant sessions live and keep their whole conversation until they are stopped.
I chose MCP so that Claude Code can call the bridge as named tools, with a permission rule for each tool and one user-level setup across projects, including sessions opened from the phone. The server’s instructions tell Claude when to use each tool, and the Python code fixes the consultant’s sandbox settings instead of leaving Claude to write the command. A fixed script could enforce the same settings, and a personal skill could also be available in every project, but neither gives Claude a set of named tools with their own permissions.
How it has grown
nouez began as a bridge for consulting Codex. Later versions added more, as my use of it grew:
- Claude no longer waits: a small background process wakes it when the reply is ready.
- A consultant started earlier in a project is reused with its context, even from a new Claude Code session.
- Claude Code sessions can be opened, listed and closed, including in a folder Claude Code does not trust yet, once I confirm that I trust it, so new projects can start from the phone.
- A consultant’s model and reasoning effort can be changed for a hard question without losing its context.
- Results now say what a reused consultant has been doing and what it runs on, and there is a tool for reading its history.
My harness and models
Claude Code is my main harness, running Claude Opus 5.5. My consultants run on gpt-6.1-sol at high reasoning effort, which is set in my Codex configuration. I find the pairing works well, because the two models are trained by different companies and read the same files differently. They are not the only choice: a consultant can use any model that Codex offers, and Claude can switch a consultant’s model or effort for a single hard question and then switch it back.
Try it
The code, with setup instructions, the tested versions and a full security note, is at github.com/jibarix/nouez under the MIT licence. It needs Claude Code, the Codex CLI and Python, and its doctor command checks the setup.
Before you enable it, know its limits. Agreement between two models does not establish truth, since both can share a gap in their knowledge. A consultant can read anything its sandbox permits, typically including files outside its project. Any Claude Code session with the server enabled can message any Codex session on the computer, and sessions I opened myself keep their own permissions and may be able to write. nouez works on one machine only, and it depends on an experimental Codex protocol that can change between releases, which is why the repository records the last combination of versions known to work.
Method and limits
The description of nouez follows its README, changelog and code at version 0.12.3. The harness-note examples compare that note’s drafts with its published text. The kinds of consultation come from the messages Claude Code sent to consultants on my laptop, sorted by an agent with keyword rules and checked by hand on a sample; about one label in ten was wrong, so the order of the kinds is approximate.
Use of AI tools
Claude (Anthropic) drafted this text under my direction, and a Codex model (OpenAI), reached through nouez, reviewed it; I checked the claims and the final wording.