MCP servers defined in config.toml are not detected/visible after update
What version of Codex is running?
codex-cli 0.88.0
What subscription do you have?
pro
Which model were you using?
5.2 codex
What platform is your computer?
Win11
What terminal emulator and version are you using (if applicable)?
VSCode/Windows Terminal
What issue are you seeing?
After updating to the latest version, Codex fails to detect or display any configured MCP servers.
Previously, my MCP servers (configured via config.toml) appeared correctly under the "Resources" or "MCP" section. Now, the "Resources" section is either missing entirely or empty.
I have verified that the config.toml file contains valid [mcp_servers] entries. When running commands to list MCP servers (e.g., via TUI /mcp or CLI), no servers are returned, as if the configuration file is being ignored or silently failing to load.
Config snippet (Sanitized):
Ini, TOML
[mcp_servers.perplexity-mcp]
type = "stdio"
command = 'C:\Users\User\.local\bin\uvx.exe'
args = ["perplexity-mcp"]
[mcp_servers.context7]
type = "stdio"
command = 'C:\Program Files\nodejs\npx.cmd'
args = ["-y", "@upstash/context7-mcp@latest"]
What steps can reproduce the bug?
Define one or more valid MCP servers in config.toml (using standard stdio configuration).
Launch Codex (either via CLI or IDE extension).
Attempt to access MCP resources (e.g., check the "Resources" tab in IDE, or type /mcp in TUI).
Result: The list is empty. No errors are displayed regarding TOML parsing, but the tools are unavailable to the model.
What is the expected behavior?
Codex should parse the existing config.toml on startup and load the defined MCP servers, making them visible in the UI and accessible to the model for tool use.
Additional information
This issue started occurring immediately after the recent update.
I have verified that the external commands (like uvx or npx) work correctly in the terminal outside of Codex.
No specific error message is printed to stderr indicating why the config is not being loaded.
6 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
This isn't a duplicate issue; it's a new problem that appeared in the new version.
Where is this config.toml file located? Is it your global config, or is it in your project folder (a project-local config)?
Can you confirm that if you revert back to version 0.87 that the problem goes away? I want to confirm that this was a regression rather than a config change.
Yes, I can confirm that reverting to version 0.87 resolved the issue immediately.
Regarding the configuration location, I actually use both:
Global: ~/.codex/config.toml
Project-Local: F:\Homework\...\.codex\config.toml (inside the workspace root)
The specific regression behavior is:
In v0.87: It correctly loads both files (merging the project-local settings with the global settings).
In the new version: It seems to ignore the project-local config entirely and only loads the global config from ~/.codex.
My missing MCP servers were defined in the project-local config, which explains why they disappeared after the update but reappeared after rolling back.
Root cause I found here: project-local
.codex/config.tomlis now intentionally only loaded for trusted projects. If the workspace isn’t trusted, that layer is skipped, so any[mcp_servers]defined there will “disappear” while the global~/.codex/config.tomlstill loads.I put together a patch (openai/codex#9708 — now closed since @etraut-openai mentioned a fix is already in progress) that adds clearer diagnostics + a simple
codex trustcommand to explicitly trust a workspace, with regression tests covering the untrusted→trusted flow.Tagging @AyinMostima for visibility.
Great, thank you very much for your work.