Step-by-step setup for local Codex projects in OpenAI's Windows app
This guide covers integrating KaKeKiKoKu with local Codex projects in the ChatGPT desktop app.
OpenAI now documents the Windows application as the ChatGPT desktop app. Codex is the local coding-agent experience inside that app: it opens repositories, runs parallel project threads, reviews changes, and connects to local MCP servers. The historical URL of this KaKeKiKoKu guide remains /codex/ so existing links continue to work.
The ChatGPT desktop app, Codex CLI, and the Codex IDE extension use the same Codex host and share config.toml. Configure KaKeKiKoKu once in that shared layer and you can move between those local surfaces without registering it again.
This guide covers the native Windows setup. OpenAI also publishes desktop clients for macOS and Linux. KaKeKiKoKu has no public macOS build; its integration with the Linux desktop client is not tested here. For terminal use on Linux, see the Codex CLI guide.
The desktop app does not support KaKeKiKoKu channels today. KaKeKiKoKu therefore works in Precise mode: Codex can use the KaKeKiKoKu MCP tools, while KaKeKiKoKu UI actions that would push messages to the AI copy text to the clipboard for you to paste.
Use a local Codex project thread. A normal ChatGPT conversation and a hosted web chat do not expose your Windows KaKeKiKoKu process. Open a local project in the desktop app so the Codex host can start the local STDIO server.
Before you start the KaKeKiKoKu integration flow, you need:
winget install --id 9PLM9XGG6VKS -s msstore↑ Back to topSecurity check. Avoid third-party download mirrors. Use OpenAI's documentation or the Microsoft Store package ID shown above.
The command Codex must start is the KaKeKiKoKu Windows alias from Quick Start Step 1:
wuniq-mcp.exeThe current app has a dedicated MCP server panel. Direct configuration remains useful when you want project scope or need to inspect the exact settings.
wuniqSTDIO or local commandwuniq-mcp.exeImportant. Use the alias
wuniq-mcp.exe. Do not point Codex at an absolute path underC:\Program Files\WindowsApps\...; Windows protects that folder and external MCP clients can hit Permission denied or EPERM.
For a user-wide server shared by the desktop app, Codex CLI, and the IDE extension, edit:
%USERPROFILE%\.codex\config.tomlCreate the file if it does not exist, then add:
[mcp_servers.wuniq]
command = "wuniq-mcp.exe"
enabled = true
startup_timeout_sec = 30If the file already contains other MCP servers, keep them. Add only the wuniq block.
↑ Back to topProject-scoped alternative. A trusted repository can keep the same block in
.codex/config.toml. Project configuration is loaded only after the repository is trusted and is useful when you do not want KaKeKiKoKu enabled in every Codex project.
This is the second of the two pieces every integration needs. Codex reads AGENTS.md files before it starts work. Put KaKeKiKoKu's startup instructions where Codex will load them for the project you open.
AGENTS.md if it does not already exist.# KaKeKiKoKu Studio (MCP)
At the start of each conversation, use the `run` tool of the `wuniq` MCP server (KaKeKiKoKu Studio, which holds the project knowledge): run `open list`, then `help`, and work as `help` explains. Run `help` again if the conversation is compacted.This file only starts the connection. Don’t copy your project’s decisions into it, or into your AI’s memory or skills: they live in KaKeKiKoKu. See Avoid competing copies of project knowledge.
Global alternative. If you want the KaKeKiKoKu startup instruction in every Codex project, create this file:
%USERPROFILE%\.codex\AGENTS.mdPut the same KaKeKiKoKu block there. Project instructions in the opened folder still apply.
↑ Back to topIf you already have startup instructions. Keep your existing content and add the KaKeKiKoKu block. Codex reads the whole file.
After adding or editing MCP configuration, use the MCP panel's Restart action or close the ChatGPT desktop app completely and reopen it. Start a fresh local project thread so both the server list and AGENTS.md are loaded again.
Run open list with the KaKeKiKoKu tool.
Inside Codex, ask:
Read the project index with KaKeKiKoKu.
Codex should use the wuniq MCP tool. If it answers from general memory, ask explicitly to use KaKeKiKoKu and re-check the startup instructions file.
You can also type /mcp in the composer to inspect the servers connected to the current Codex host.
KaKeKiKoKu runs as a local web app, and the ChatGPT desktop app gives Codex a built-in Browser that both you and the AI can use. Open KaKeKiKoKu there and your knowledge workspace sits beside the conversation: one window, one screen, no switching back and forth.
That changes what you can see while Codex works. The KaKeKiKoKu UI blinks green when the AI reads knowledge and red when it writes, while the connection indicator shows whether Codex is attached. In a separate window you miss that because you are watching the conversation. Inside Codex, you watch your knowledge layer being used as it happens.
This step is optional. KaKeKiKoKu works exactly the same without it.
KaKeKiKoKu must already be running. Codex attaches its Browser to the live Knowledge Engine; it never starts another one.
Check that the Browser capability is available in the current desktop task. OpenAI documents it as a built-in desktop browser; do not assume that a plugin named Browser must be installed. This pane is not available in Codex CLI or the IDE extension. You can always use KaKeKiKoKu in your normal browser.
Open a local Codex project in your working directory and paste this:
Set up KaKeKiKoKu in Codex's built-in Browser for my working directory.
1. Check that the Browser capability is available in this desktop task. If
it is not, tell me and keep using KaKeKiKoKu in my normal browser. Do not
assume a particular plugin must be installed.
2. Find the port. Try http://localhost:5050 first, which is the KaKeKiKoKu
default. If nothing answers there, read the "KnowledgeEnginePort"
value from %LOCALAPPDATA%\Wuniq\app-settings.json. If you still cannot
determine it, ask me: the port is visible in the address bar of the
KaKeKiKoKu window.
3. Confirm KaKeKiKoKu is actually responding on that port. If it is not, stop
here and tell me to start KaKeKiKoKu first. Do not try to start KaKeKiKoKu
yourself. The browser must attach to the existing Knowledge Engine;
this step only displays the application that I have already started.
4. Create AGENTS.md in my working directory if it does not exist. Add the
following instruction, keeping all existing content and merging it into
an existing KaKeKiKoKu section instead of duplicating it. Replace PORT with
the port you found in step 2 before writing it:
At the start of every task, use Codex's built-in Browser to open KaKeKiKoKu
at http://localhost:PORT/. Reuse an existing KaKeKiKoKu tab instead of
opening a duplicate, and make the Browser visible. If Browser is
unavailable or the page does not load, tell me in one line and carry
on. Do not try to start KaKeKiKoKu yourself.
5. Now use the built-in Browser to open KaKeKiKoKu, make it visible, and
confirm the KaKeKiKoKu UI loaded. Reuse an existing KaKeKiKoKu tab if one is
already open.
6. Tell me exactly which file you created or changed.Codex changes one file: AGENTS.md. Unlike Claude Code Desktop, Codex does not need a browser launch configuration — the instruction itself is enough to open or reclaim the local KaKeKiKoKu tab in later tasks. Remove that paragraph and Codex stops opening it automatically.
Why the port is not hard-coded. 5050 is the default, but KaKeKiKoKu lets you change its local port. The prompt discovers the active value before it writes the durable instruction. If you move KaKeKiKoKu later, run the prompt again.
↑ Back to topThe first run may ask for access. Approve
localhostwhen the Browser asks. Browser permissions belong to Codex, not KaKeKiKoKu, and you can review them later under Settings > Browser.
Use the app's controls to choose model, reasoning effort, and permission mode. KaKeKiKoKu does not require full access or a special model. It only needs Codex to be allowed to start and call the local KaKeKiKoKu MCP server.
For normal KaKeKiKoKu use, start with a local project, the model Codex recommends for your account, and Ask for approval beneath the composer. Increase autonomy only when you want Codex to edit or run commands more freely.
The Windows app can run its agent natively in PowerShell or inside WSL2. This guide configures the native Windows host and the Windows command wuniq-mcp.exe. A WSL-hosted Codex process has a different home directory and cannot assume the native Windows MCP command is available.
Models and plan limits change over time. Use OpenAI's official Windows app documentation and the model selector in the app for the current state.
↑ Back to topIn Precise mode, KaKeKiKoKu cannot push messages to Codex. This changes the daily flow:
sync when you want it to reconcile code and context.↑ Back to topRecommended KaKeKiKoKu configuration. Go to KaKeKiKoKu > Settings > Application > AI Sync and choose Precise mode.
wuniq-mcp.exe.config.toml, verify the block is under [mcp_servers.wuniq].config.toml edited by hand — check the encodingIf you used Option B and Codex does not load KaKeKiKoKu, make sure config.toml was saved as UTF-8 without BOM. Some editors — classic Notepad's UTF-8 option, or PowerShell's Set-Content -Encoding UTF8 — prepend a hidden 3-byte marker (a BOM) at the start of the file that strict configuration parsers reject. The marker is invisible in the editor; switch the file's encoding to UTF-8 (without BOM) and save again.
Most often this means the MCP entry points directly into C:\Program Files\WindowsApps\.... Replace that path with the alias:
wuniq-mcp.exewhere.exe wuniq-mcp.exe returns nothingWindows cannot find the KaKeKiKoKu MCP alias. Check that KaKeKiKoKu is installed, then close and reopen PowerShell or Codex. The alias normally lives through Windows' app execution alias path under %LOCALAPPDATA%\Microsoft\WindowsApps.
AGENTS.md exists in the project or global Codex home.Check the capabilities available to this local desktop task and the current Browser documentation. Availability can depend on the installed app and workspace policy. Use your normal browser if this task has no browser capability. The MCP connection remains usable without the pane; Codex CLI and the IDE extension do not provide this built-in desktop pane.
KaKeKiKoKu is not running yet, or the port stored in AGENTS.md no longer matches KaKeKiKoKu's current local port. Start KaKeKiKoKu yourself, give a fresh installation a few seconds, and run the setup prompt again. Do not ask Codex to launch the installed executable: KaKeKiKoKu and the AI client deliberately start independently, and the Windows Store package lives behind restrictive WindowsApps permissions.
Make sure the instruction in AGENTS.md says to reuse an existing KaKeKiKoKu tab. Close the duplicates once, keep one KaKeKiKoKu tab open, and start a fresh task so Codex reloads the corrected instruction.
Open a local Codex project thread in the desktop app. Hosted or ordinary chat surfaces do not read the local Codex config.toml in the same way and cannot reach the KaKeKiKoKu process on your PC.
The configuration on this page is for the native Windows host. If you switch the app to WSL, restart the app as OpenAI requires and treat that Codex host as Linux: it has a separate ~/.codex, PATH, and process environment. Do not copy wuniq-mcp.exe into a WSL configuration and assume the bridge is validated.
Email support@kakekikoku.com with your operating system, Codex version, and the MCP configuration you are using. KaKeKiKoKu can also generate a diagnostic report from the Support function in the app.
↑ Back to top