MCP servers not detected in Codex VS Code extension (but working in Codex CLI)

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #6215
• #5929

Powered by Codex Action

Just to clarify, this is not a duplicate of #5929 or #6215.
Both of those refer to older Codex releases and .vscode/mcp.json integration issues on macOS.

My report (#6465) is about Codex VS Code extension (v0.56.0) failing to detect MCP servers already working in Codex CLI, even though both share the same ~/.codex/config.toml configuration.

Reproduction shows CLI detects Playwright/Supabase MCP successfully, while list_mcp_resources in IDE remains empty — indicating the VS Code extension isn't spawning or connecting to the same MCP runtime session.

Tested on Windows 11 + Node v22, so this is cross-platform and not related to .vscode/mcp.json.

Please keep this open — it's a session isolation bug between Codex CLI and IDE, not a config syntax issue.

I have the same problem:
codex-cli 0.57.0
Codex VScode extension: 0.4.38

My ~/.codex/config.toml:

model = "gpt-5-codex"
model_reasoning_effort = "medium"

[projects."/path/to/back-end"]
trust_level = "trusted"

[mcp_servers.postgres]
command = "/usr/local/bin/toolbox"
args = ["--prebuilt","postgres","--stdio"]
env = {"POSTGRES_HOST"= "localhost", "POSTGRES_PORT"= "43256", "POSTGRES_DATABASE"= "my_db_name", "POSTGRES_USER"= "my_db_usernamename","POSTGRES_PASSWORD"= "my_db_pass"}

Codex CLI can access the MCP server and can list the tables of the DB.
While Codex VScode extension reports this:

resources/list failed: unknown MCP server 'postgres'

Can confirm i have the same issue on a fresh install Win 11 pro.

CLI v0.57.0
VSC extention v0.4.39

Using CLI i can access larvel boost and use the tools, however i cannot through the extention.

Also not working on Mac.

CLI v0.58.0
VSC extension 0.4.40

When prompting the chat to check the status of the tools, it's thinking always outputs:

MCP client for figma failed to start: handshaking with MCP server failed: expect initialized result, but received: Some(EmptyResult(EmptyObject))

I have the same problem:

CLI v0.57.0
Extensão VSC v0.4.39

Linux

I also have the same problem:

CLI v0.58.0
Extension v0.4.40

My config.toml:

model = "gpt-5.1-codex"
model_reasoning_effort = "medium"

[mcp_servers.svelte]
command = "npx"
args = ["-y", "@sveltejs/mcp"]
startup_timeout_sec = 30

Using Windows with WSL. CLI works fine, the extension times out.

I'm having the same issue in Win11 and in WSL.

Extension v0.4.43
CLI v0.58.0

I'm having the same issue in Win11, WSL not tested.

Extension v0.4.43
CLI v0.58.0

I'm having the same issue in VS Code Insiders Linux
Extension 0.5.46
CLI v0.60.1

.codex/config.toml:

model = "gpt-5.1-codex-max"
model_reasoning_effort = "high"

[mcp_servers.laravel-boost]
command = "php"
args = ["artisan", "boost:mcp"]

chat:

what mcp that currently installed that you can access?

I don’t see any MCP servers/resources exposed in this environment right now—both the resource and resource-template lists came back empty. If you expected some to be available, let me know their names or restart/connect them so I can access them.

but codex recognized installed mcp in CLI

Same here on MacOS. Any workaround / fixes?

I am having the same issue on Win 11/ WSL
Extension: v0.4.46
Codex-CLI 0.63.0

Can get the CLI to utilize the MCP but not the Codex extension

same, it stopped recognizing the .vscode\mcp.json mcp servers per project. i'm in MacOS

This happened to me "randomly" recently as well (MCP servers don't work). Made a thread about it with more details here. I ended up cancelling my ChatGPT sub as Codex was the main readon I was keeping it. But subscribed to this thread just in case it gets addresssed and I decide I want ChatGPT again.

Thanks for your feedback all, let's try to get to the bottom of this:

For starters:

What mcp that is currently installed that you can access?

does not seem to be a reliable way to ask the model what is currently active (despite it being a reasonable prompt)

What mcp__ prefixed tools do you have access to?

Could somebody run the above query to verify if there are issues or if the model was just confused by the way you were asking before.

We are also working on a proper MCP configuration UI in the extension that will also make this much more clear.

hi @gpeal that's always the way to ask the model, you even ask only "can you detect mcp tools?" and he would detect them. It stopped detecting them, i don't remember which was the version from codex, but

<img width="759" height="341" alt="Image" src="https://github.com/user-attachments/assets/29b7427e-9a92-4f12-b859-87a80b93f27e" />

<img width="518" height="550" alt="Image" src="https://github.com/user-attachments/assets/afb0965b-e606-4382-982d-4cd3d1b9ef97" />

.vscode/mcp.json

{
"servers": {
"herd": {
"command": "php",
"args": ["/Applications/Herd.app/Contents/Resources/herd-mcp.phar"],
"env": {
"SITE_PATH": "/Users/brunogomes/Documents/Herd/projectx"
}
},
"shadcn": {
"command": "npx",
"args": ["shadcn@latest", "mcp"]
},
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest"]
},
"sequential-thinking": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
},
"chrome-devtools": {
"command": "npx",
"args": [
"chrome-devtools-mcp@latest",
"--channel=canary",
"--headless=true",
"--isolated=true"
]
},
"laravel-boost": {
"command": "php",
"args": ["artisan", "boost:mcp"]
},
"stripe": {
"type": "http",
"url": "https://mcp.stripe.com"
}
}
}

This is copilot, which works always, never had the issues of the MCP that we sometimes have with Codex.

<img width="607" height="678" alt="Image" src="https://github.com/user-attachments/assets/52a2e5b8-4faa-4cea-84f6-80a84965e610" />

hope you guys fix this in the next release because not having up to date documentations are not a good way for the model to work.

@Zokor Could you try with the prompt I suggested? Different models and harnesses may behave differently. I agree that Codex _should_ have a reasonable response to your message but it does seem to work more reliably with:

What mcp__ prefixed tools do you have access to?

and knowing whether it's a model behavior issue or a harness issue would help here.

@gpeal I can confirm that its not a model issue, ive asked it to use or
list what MCP tools it has and every time it will say that Laravel boost is
installed but the MCP server isnt running and it cannot see, access or use
the tools. if you do the same thing on CLI it just works.

output when running you're prompt.

I don’t see any mcp__‑prefixed tools exposed to me in this session.

Right now, the only tools I can actually call are the generic ones you see
documented (e.g. shell_command, list_mcp_resources, read_mcp_resource,
apply_patch, etc., all under the functions namespace). If your IDE shows
additional Laravel Boost MCP tools (like mcp_laravel_boost__...), they
aren’t currently wired through to this assistant run, so I can’t enumerate
or invoke them directly.

also just a FYI, Claude works with the MCP (VSCode extension also)

On Mon, Dec 1, 2025 at 3:15 AM Gabriel Peal @.***>
wrote:

gpeal left a comment (openai/codex#6465) <https://github.com/openai/codex/issues/6465#issuecomment-3594350003> @Zokor <https://github.com/Zokor> Could you try with the prompt I suggested? Different models and harnesses may behave differently. I agree that Codex should have a reasonable response to your message but it does seem to work more reliably with: What mcp__ prefixed tools do you have access to? and knowing whether it's a model behavior issue or a harness issue would help here. — Reply to this email directly, view it on GitHub <https://github.com/openai/codex/issues/6465#issuecomment-3594350003>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2VXQNDBFPRK4WUZCGINUID37OXFVAVCNFSM6AAAAACLWBYTVOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKOJUGM2TAMBQGM> . You are receiving this because you commented.Message ID: @.***>

@MatthewGranfield And it works from the CLI? Does it ever work with the extension? Does it matter if you use WSL or not?

I use WSL and VSCode in WSL mode. CLI works fine for me, loads up the MCP
tools at the start and /MCP list shows them. If I run a test prompt it sees
and uses the tools fine.

Never worked for me in the extension, but i've only tried to use it in the last 6w maybe?

also - ive never tried to use the CLI version outside of WSL. Ive tried to use the VSC version outside of WSL but i didn't get it working either.

On Mon, 1 Dec 2025, 07:04 Gabriel Peal, @.***> wrote:

gpeal left a comment (openai/codex#6465) <https://github.com/openai/codex/issues/6465#issuecomment-3594895043> @MatthewGranfield <https://github.com/MatthewGranfield> And it works from the CLI? Does it ever work with the extension? Does it matter if you use WSL or not? — Reply to this email directly, view it on GitHub <https://github.com/openai/codex/issues/6465#issuecomment-3594895043>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2VXQNEKL67G2YDNOMJMIN337PR7ZAVCNFSM6AAAAACLWBYTVOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKOJUHA4TKMBUGM> . You are receiving this because you were mentioned.Message ID: @.***>

It worked for me with the extension some versions ago :)

I don't use the CLI directly because that's wasting time with another window open getting out of the working context. So having the extension working properly is very important for the product to be viable.

Anyway here's the answer to the question:

<img width="599" height="596" alt="Image" src="https://github.com/user-attachments/assets/bece05c1-68d6-49dd-8dc1-0b2be793fd64" />

Observing with MacOS setup as well - works correctly in CLI, no MCPs in IDE.

i had the same issue but i think it works now i just askes codex extension to check my config.toml because i already setup supabase mcp and it said it can see the supabase entry and it tried to query data and it went through hope this helps

i have mac os btw

I think you may figure out the cause by running the internal codex binary directly. If you are using Unix-like OS, it is located ~/.vscode/extensions/openai.chatgpt-<version>/bin/<arch>/codex.

For me, it shows errors like:

⚠ MCP client for `<mcp-server-name>` failed to start: MCP startup failed: missing field `type`

which I solved in #7417.

Observing with MacOS setup as well - works correctly in CLI, no MCPs in IDE.

I solved this for my setup with VS Code MCP settings.

Codex in the IDE for some reason picks up these MCPs, not the ones from ~/.codex/config.toml.

I am observing something strange. The Codex doesn't regard the configured mcp server as something different.

<img width="549" height="463" alt="Image" src="https://github.com/user-attachments/assets/822e5098-38dd-4c7a-b614-d6840c76ec72" />

Note that I am using the binary I prepared myself because the default version for the VS Code addon has some bug as I mentioned in the comment from two messages ago.

I am having the same issue on Win 11/ WSL Extension: v0.4.46 Codex-CLI 0.63.0 Can get the CLI to utilize the MCP but not the Codex extension

• > My Codex extension can now detect MCP servers. I have made no changes and the extension version is the same. Issue resolved for me.

Observing with MacOS setup as well - works correctly in CLI, no MCPs in IDE.

How did you verify whether there are no MCPs in the IDE?

I am observing something strange. The Codex doesn't regard the configured mcp server as something different. <img alt="Image" width="549" height="463" src="https://private-user-images.githubusercontent.com/61399424/521338285-822e5098-38dd-4c7a-b614-d6840c76ec72.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.ghV_HbzMWcmJQk0X4EfBCtTGnFep6FjAHnliWAMcfcA"> Note that I am using the binary I prepared myself because the default version for the VS Code addon has some bug as I mentioned in the comment from two messages ago.

Please check my previous comment with regards to model behavior and how to ask here. It's not ideal but worth noting.

Having the same issue with Codex VS Code Extension not recognizing the mcps. "None detected at the moment. The list_mcp_resources and list_mcp_resource_templates calls both return empty, so no mcp__* tools are loaded."

Problem persists with the 0.64.0 release:

<img width="2176" height="750" alt="Image" src="https://github.com/user-attachments/assets/6a376ffd-3b66-48f7-a5a7-e1776a2161ab" />

Just a quick update, if i try and run the extention's codex file from the term i get:

~/.vscode-server/extensions/openai.chatgpt-0.4.47/bin/linux-aarch64$
⚠ MCP client for `laravel-boost` failed to start: MCP startup failed: No such file or directory (os error 2)

⚠ MCP startup incomplete (failed: laravel-boost)

mcp still not working in codex extension - 7th january 2026

@PangeranSilaen

Try this:

I think you may figure out the cause by running the internal codex binary directly. If you are using Unix-like OS, it is located ~/.vscode/extensions/openai.chatgpt-<version>/bin/<arch>/codex. For me, it shows errors like: `` ⚠ MCP client for <mcp-server-name> failed to start: MCP startup failed: missing field type `` which I solved in #7417.

If you have an error, it should be reported as another issue. If you don't, I think it's truly a problem of this extension.

@linuxmetel the issue still persists, no MCP's being detected

@Zokor according to this comment, just asking if it has any access to an mcp server is not enough. So I don't know what happen until you send another screenshot or suchlike. And if you can, also try ~/.vscode/extensions/openai.chatgpt-<version>/bin/<arch>/codex.

i've created another, but it's a lot of time in the "dark" with mcp issues

(https://github.com/openai/codex/issues/8836)

@Zokor according to this comment, just asking if it has any access to an mcp server is not enough. So I don't know what happen until you send another screenshot or suchlike. And if you can, also try ~/.vscode/extensions/openai.chatgpt-<version>/bin/<arch>/codex.

@linuxmetel these screenshots? :o

<img width="789" height="427" alt="Image" src="https://github.com/user-attachments/assets/4c1102bd-56af-4bc8-ad76-6219ed42cc93" />

<img width="872" height="196" alt="Image" src="https://github.com/user-attachments/assets/09c72e3c-11b1-425e-a8b7-ce9ad0813fdf" />

@Zokor the codex-cli binary the extension uses is there. I meant the screenshot when you runned it. In order to make it clear if it's the problem of codex-cli binary or not.

And looking at your issue, I don't know what is the problem. Doesn't the extension detect mcp servers you added in config.toml?

Codex was detecting the setup of the .vscode/mcp.json without using the toml some weeks ago.

<img width="592" height="573" alt="Image" src="https://github.com/user-attachments/assets/14413adc-605b-4c62-996b-bc1f9d4de56a" />

i went back on it and the config.toml is working again, thank you. i'll start using it instead of the .vscode\mcp.json. but it would be nice if we start having standards for these things :) cheers

@Zokor, the Codex IDE extension has never supported .vscode/mcp.json. It reads its MCP configuration from config.toml. In recent versions of the IDE extension, we added a dedicated UI for configuring MCP servers. You can access this through the gear icon in the upper right of the panel. Select "Open MCP Settings".

@etraut-openai hi, thank you for the reply, that's strange because, it's what i used since i made my first subscription for a full month between October and November only the .vscode/mcp.json and it was working. ( the config.toml wasn't working ) was cool because i was also using copilot initially and then i switched to Codex in October, plus it was a way i had to commit per project the mcp's i was using

same issue

From codex in VSCode terminal
do you see nx mcp?
• Yes — Nx MCP tools are available (e.g., nx_workspace, nx_project_details, nx_docs, etc.).

From Codex Extension
`do you see nxmcp

I don’t see Nx MCP tools (like nx_workspace, nx_project_details, nx_docs) in the current session tool list, so I can’t use them right now. If you expect them to be available, we can check the MCP server configuration—just say the word and I’ll request approval to read config.toml (outside the workspace).`

Config.toml
[mcp_servers.nxmcp]
command = "npx"
args = ["nx", "mcp"]

It seems that if I start a new conversation in VSCode extension, it does see the right MCP servers / functions. But if I switch right back to the previous conversation, it doesn't see it

Maybe you can try using this approach: Workaround: (https://gist.github.com/DevloperHS/0bd80665579d80fccac163192c5b6448#file-steps)

Here is proof it works:

<img width="948" height="714" alt="Image" src="https://github.com/user-attachments/assets/d175acbe-c2e6-4640-875c-2e2e59f8b2ee" />

To fix this, I would like to request a new feature, which would allow CODEX_HOME to be set in the ./.vscode/settings.json. Doing so allows us to have 100% repo local Codex, which even allows for using separate work and a personal accounts for separate repos. It should also be easy to implement, since all that's required is setting the CODEX_HOME environment variable before calling codex. This works great when using the CLI, and it even works in VSCode if you launch VSCode with CODEX_HOME set, but that is a huge pain to do every time you need to open your project.

I have a similar issue, but only for project-level config.toml. VS Code Codex chat does not pick up project-level MCP servers, while the Codex CLI does. User-level config works fine.

this is piece of shit, settings doesn't able to save changed value, how thfcuk this dev asking 200$ in month but can't make proper UI for his own product. using cli to add mcp is another headache

This is not related to windows-os btw! Incorrect tag!

Is there really no update on this after months? Or even a hint of what causes it? Is this no affecting all users? How are people even able to work without MCPs?

I was able to use MCPs just fine in the extension till a few days ago, when I installed the Codex app on Windows, and now it went haywire. The extension won't be able to use any of the custom MCPs (even after uninstalling the Codex app, which was even worse).

Here's the bug report it created for me after investigation:

Custom MCP servers from ~/.codex/config.toml load and work in standalone codex v0.132.0 (/mcp shows homeassistant + wordpress MCPs and tools execute), but the VS Code ChatGPT extension runtime only exposes curated MCP namespaces (codex_apps/context7) to the assistant tools. This worked a few days ago. 

Environment:
- WSL Ubuntu 24.04
- Extension: openai.chatgpt-26.513.21555-win32-x64
- codex app-server launched by extension
- ~/.codex/config.toml contains custom mcp_servers entries

I took another pass through this thread and the current reports, and it may help to split this into a small diagnostic matrix. Right now several distinct failure modes all present as “the extension has no MCP tools”.

Suggested triage buckets:

1. Config not loaded

• ~/.codex/config.toml works in CLI, but the extension does not read it.
• Project-level config works in CLI but not the extension.
• CODEX_HOME differs between CLI, VS Code, WSL, remote SSH, and the standalone app.
• The extension MCP settings UI shows a different source of truth than the CLI.

1. Config loaded, server not started

• Extension can see the MCP entry, but startup fails because PATH/env/cwd differs from the user’s terminal.
• Examples from this thread include npx, PHP/Laravel, WSL, and extension-bundled codex binary differences.
• A visible startup-status surface would help distinguish this from “no config”.

1. Server started, tools not exposed to the model/session

• Recent comments suggest standalone CLI /mcp can show tools while the VS Code assistant only exposes curated namespaces or old conversations miss tools that new conversations can see.
• That sounds different from a plain config parse or process-start failure.

A narrow MVP would be a “MCP diagnostics” view or command in the extension that prints, for the current chat: effective CODEX_HOME, config files considered, server startup status, and count/list of exposed mcp__* tools. That would give reporters a stable artifact instead of relying on prompts like “what MCPs do you see?”, which maintainers have already noted can be model-behavior dependent.

Per docs/contributing.md, I’m sharing analysis rather than opening an unsolicited PR. If useful, this could be narrowed into a test matrix across local VS Code, WSL, remote SSH, and project-level config.

Could someone from the core team take a look at this issue? I'm eager to switch over to Codex, but this problem is making the transition quite difficult.

This is more relevant now than ever during the great GHCP exodus.... Looking forward to an update (and confirmed fix) for this soon!

I added a small evidence packer for this MCP discovery/config-scope mismatch class in trace-to-skill@0.1.64:

npx trace-to-skill@0.1.64 demo mcp-discovery-mismatch
npx trace-to-skill@0.1.64 codex-report ./runs --output openai-codex-mcp-discovery.md

It is not a fix; it just helps reports include consistent fields that seem useful for this thread: app/CLI/extension version, OS/IDE/remote or WSL state, workspace root, effective CODEX_HOME, config files considered (~/.codex/config.toml, project .codex/config.toml, IDE-level MCP config), redacted MCP sections, trust/profile/default-permissions state, codex mcp list/get output, CLI-vs-Desktop/VS Code comparison, loaded config path/log lines, whether moving the same server to user-global config fixes it, restart/reload/new-conversation results, and whether the current session exposes any mcp__* tools.

Release: https://github.com/grnbtqdbyx-create/trace-to-skill/releases/tag/v0.1.64

@grnbtqdbyx-create, please stop spamming our issue tracker with AI-generated responses like this. If you continue, I will ask github to ban your account.

Same issue here, it's incredibly frustrating. I'll go back to Claude in the meantime.

Same issue here. Win 10. VS Code Version: 1.123.0
I wish we'd at least get some better error indications as I'm not entirely sure what exactly is failing.

Right now there's ZERO indication in VSCode Codex Extension whether MCPs work at all. I tried adding MCPs via codex.toml and via 'GUI' (quoted, because GUI is just a vibe-slopped low-effort attempt to represent toml-strings, withiot any extra functionality compared to direct toml editing). Ample errors in logs though:

2026-06-10 03:32:43.652 [warning] [CodexMcpConnection] cli: message="codex_mcp::connection_manager: Failed to list resource templates for MCP server 'context7': Mcp error: -32601: Method not found"
2026-06-10 03:32:43.652 [warning] [CodexMcpConnection] cli: message="codex_mcp::connection_manager: Failed to list resource templates for MCP server 'resend': Mcp error: -32601: Method not found"
2026-06-10 03:32:43.689 [warning] [CodexMcpConnection] cli: message="codex_mcp::connection_manager: Failed to list resources for MCP server 'Supabase-local': Mcp error: -32601: Method not found"
2026-06-10 03:32:43.710 [warning] [CodexMcpConnection] cli: message="codex_mcp::connection_manager: Failed to list resource templates for MCP server 'Supabase-local': Mcp error: -32601: Method not found"

Was this thing supposed to replace Copilot at some point? Very funny.

FWIW, I'm using VSCode Remote.