Codex used to run as an MCP server for Claude Code, but the codex mcp-server command was removed from the CLI. This article shows what broke, which replacements work today, the exact install commands, sandbox settings, and fixes for common errors.
If you typed codex mcp-server into a terminal and got Error: stdin is not a terminal, your machine is fine. The command is gone. For a while, Codex MCP for Claude Code meant one setup: run Codex as an MCP server, register it with Claude Code, and let one coding agent ask the other for a second opinion. OpenAI has since removed that subcommand, and every tutorial that still shows it leads to a dead end.
This article explains what changed, which replacements work today, and the exact commands for each one. You get the official plugin, a community MCP wrapper, a short-lived stopgap with an older CLI, and the reverse setup where Codex calls Claude Code. Each command comes from OpenAI's documentation, the plugin repository, or the wrapper's README, so you can paste it without guessing.
💡 Short version: install the official Codex plugin for Claude Code. Choose a community MCP wrapper only when you need Codex to show up as real MCP tools.
Codex CLI used to ship a subcommand, codex mcp-server, that started Codex as a stdio MCP server. Claude Code could launch it like any other server, so a single claude mcp add line turned Codex into a tool. The server exposed two tools:
| Tool | What it did | Main inputs |
|---|---|---|
codex | Started a new Codex session | prompt, approval-policy, sandbox |
codex-reply | Continued an existing session | prompt, threadId |
The first call returned a threadId. Passing it to codex-reply kept the conversation going, so Claude Code could hand Codex a task, read the answer, then follow up in the same session. The approval-policy field accepted untrusted, on-request or never, and sandbox accepted read-only, workspace-write or danger-full-access.
The old registration looked like the line below. It no longer works on a current CLI:
claude mcp add codex -- codex mcp-server
OpenAI's MCP documentation now states that the codex mcp-server command and the standalone codex-mcp-server binary have been removed. On a current CLI, mcp-server is not a recognized subcommand, so the CLI treats the word as a prompt and tries to open its interactive interface. Claude Code launches servers over a pipe with no terminal attached, so the CLI stops with Error: stdin is not a terminal. A bug report in the Quest project traces the failure to exactly this fall-through.
Inside Claude Code the symptom is a server that shows as failed in /mcp, with no tools listed. Nothing in your config is wrong. The thing it points at no longer exists.
A migration write-up dates the deprecation notice to Codex CLI 0.149.1, released on August 24, 2026. Bug reports from several projects that depended on the command, including Quest and the second-opinion skill from Trail of Bits, name Codex CLI 0.154.0 as the release where it disappeared. Check what you run:
codex --version
Any version from 0.154.0 onward lacks the subcommand, so no amount of config editing will bring it back.
Pick based on how you want Codex to appear inside Claude Code. The table shows the trade in one glance:
| Option | Maintained by | Appears as | Catch |
|---|---|---|---|
| Codex plugin for Claude Code | OpenAI | Slash commands such as /codex:review | Not MCP tools |
| Community MCP wrapper | A third party | MCP tools such as ask-codex | Unofficial, wraps codex exec |
| Older Codex CLI | OpenAI, frozen | The original codex and codex-reply tools | No updates, breaks on upgrade |
If your goal is a second opinion on code, the plugin is the shortest path. If you already have prompts, scripts or subagents that call MCP tools by name, the wrapper keeps that shape with the least rewriting.
OpenAI publishes openai/codex-plugin-cc. It wraps the Codex app server through the codex binary already on your machine, so it reuses your sign-in and your Codex configuration. It needs Node.js 18.18 or later plus a ChatGPT subscription (the Free tier counts) or an OpenAI API token. The trade: you get slash commands and background jobs, not MCP tools.
@cexll/codex-mcp-server keeps the MCP shape. It calls codex exec under the hood and exposes the tools ask-codex, brainstorm, ping and help. It needs Node.js 18 or later and a Codex CLI that is installed and signed in. Because it is third-party code that runs against your repository, read it before you trust it, and expect it to trail Codex CLI changes by days or weeks.
You can install a Codex CLI release from before the removal and keep codex mcp-server alive:
npm install -g @openai/codex@<a-version-before-0.154.0>
Confirm the subcommand exists before you wire it in. This only buys time. You freeze out new models and fixes, and the next upgrade breaks the setup again.
💡 Treat the older CLI as a bridge for a week, not as an architecture.
Install the Codex CLI first. The plugin relies on that global binary, and it uses whatever sign-in the binary already holds.
From a normal shell, install the CLI:
npm install -g @openai/codex
Then run these inside Claude Code, one at a time:
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup
/codex:setup checks that Codex is installed and authenticated. If it reports a problem, fix that before anything else, because every other command depends on it.
The plugin adds a small family of /codex: commands:
| Command | Use it for |
|---|---|
/codex:review | A standard, read-only review of your changes |
/codex:adversarial-review | A steerable review that challenges design choices |
/codex:rescue | Handing a stuck task to Codex through a subagent |
/codex:transfer | Creating a persistent Codex thread from the current session |
/codex:status | Listing running and recent Codex jobs |
/codex:result | Showing the final output of a finished job |
/codex:cancel | Stopping an active background job |
/codex:setup | Verifying installation and sign-in |
A habit that works: after Claude Code finishes a change, run /codex:review to get a second model's opinion before you commit. Reach for /codex:rescue when Claude Code loops on the same bug, and watch /codex:status while the job runs in the background.
Choose this route when you need real MCP tools. The wrapper sits between Claude Code and the Codex CLI, translating tool calls into codex exec runs.
npm install -g @openai/codex
claude mcp add codex-cli -- npx -y @cexll/codex-mcp-server
Then type /mcp inside Claude Code. The server should appear as connected, with its tools listed underneath. claude mcp list shows the same from the shell.
Add --scope project to the claude mcp add command and Claude Code writes the server into .mcp.json at the repository root, so the whole team gets the same setup:
{
"mcpServers": {
"codex-cli": {
"command": "npx",
"args": ["-y", "@cexll/codex-mcp-server"]
}
}
}
Claude Code asks each teammate to approve a project server before it runs, which is the right behavior for a tool that can touch files.
The wrapper passes Codex's safety controls through, so you decide how much freedom Codex gets. The README lists these options:
| Setting | Options |
|---|---|
| Sandbox | read-only, workspace-write, danger-full-access |
| Approval policy | never, on-request, on-failure, untrusted |
💡 Start with
read-onlyfor reviews and questions. Move toworkspace-writeonly when you want Codex to edit files. Pairingdanger-full-accesswithneverremoves every brake, so keep that combination out of shared repositories.
The README also notes smart defaults since version 1.2: the wrapper picks workspace-write when a task needs it, to prevent permission errors. Check which mode a run actually used before you assume it stayed read-only.
The connection works in both directions. Claude Code can call Codex, and Codex can call Claude Code or any other MCP server you trust.
Codex consumes MCP servers natively. Add one from the CLI:
codex mcp add my-server -- npx -y your-mcp-server-package
Or write it into config.toml as an [mcp_servers.<name>] table:
[mcp_servers.my-server]
command = "npx"
args = ["-y", "your-mcp-server-package"]
startup_timeout_sec = 20
tool_timeout_sec = 120
Stdio servers also accept env, cwd, enabled_tools and disabled_tools, which lets you hide tools Codex should never call. HTTP servers take a url plus authentication options.
Claude Code can serve its own tools over MCP with claude mcp serve. Register that command in Codex:
codex mcp add claude-code -- claude mcp serve
Codex then sees Claude Code's tools, such as file reads, edits and shell access, as MCP tools. Try it in a throwaway repository first, because two agents with write access can overwrite each other's edits.
Most failures fall into five buckets. Match the symptom, then apply the fix:
| Symptom | Likely cause | Fix |
|---|---|---|
Error: stdin is not a terminal | The mcp-server subcommand was removed | Switch to the plugin or the wrapper |
Server shows as failed in /mcp | Codex CLI missing from PATH or not signed in | Run codex --version, sign in, restart Claude Code |
/codex: commands missing | Plugin not reloaded | Run /reload-plugins, then /codex:setup |
npx fails on native Windows | npx is a .cmd shim | Register with cmd /c npx -y @cexll/codex-mcp-server |
| Server times out on first start | npx downloads the package on the first run | Install it globally, or raise MCP_TIMEOUT |
A 30-second health check catches nearly everything else:
codex --version to confirm the CLI is installed and recent.claude mcp list to confirm the server is registered and connected.ping (wrapper) or run /codex:setup (plugin).If all three pass and a task still fails, the problem is the task, not the plumbing. Narrow the prompt, drop the sandbox to read-only, and run it again.
When a setup still fails, a strong coding model can read the error with you. GPT 5.6 Sol on PicassoIA is built for coding, technical writing and multi-step reasoning, and it reads images, so a screenshot of the failing terminal works too.
codex --version output and the command that failed into the prompt field.high for tangled issues. The default, none, favors speed.low when you only want the fixed command, and high when you want the reasoning spelled out.Here is how the settings map to this kind of job:
| Setting | Values | Good for |
|---|---|---|
reasoning_effort | none, low, medium, high, xhigh | none for quick lookups, high for config conflicts |
verbosity | low, medium, high | low when you need only the command |
max_completion_tokens | Integer | Raise it together with the effort |
system_prompt | Text | Setting a role and a rule about guessing |
image_input | List of images | Terminal screenshots and diagrams |
To compare answers, run the same prompt on Claude Sonnet 5 or Claude Fable 5, both in the same Large Language Models category. Two models disagreeing about your config is a useful signal to read the docs yourself.
Setup articles and READMEs read better with real visuals: a header photo, a diagram of how two agents connect, a short clip for a release note. PicassoIA also has its own developer API and MCP connector, so the same Claude Code session can request pictures without leaving the terminal.
The API is Replicate-style: the base URL is https://api.picassoia.com/v1, requests carry a Bearer token that starts with pia_sk_, and jobs are asynchronous, so you create a prediction, poll it, then fetch the result. An account can run 5 concurrent predictions, shared across tokens and MCP connections. Four models are available through the API and the MCP connector:
Check the pricing page for what your plan includes, since API and MCP access depend on it.
Pick one section of this article, write a 50-word photo prompt for it, and generate your first image on Picasso IA. Then feed that image to a video model and turn it into a five-second clip. One afternoon of experiments will show you which model fits your docs, and the prompt you end up with is worth saving for the next article.
