Codex Desktop sometimes does not inject configured project MCP tools into a thread
Open 💬 2 comments Opened May 23, 2026 by jarvanstack
What version of the Codex App are you using (From “About Codex” dialog)?
Codex Desktop / CLI: 0.133.0-alpha.1
What subscription do you have?
pro
What platform is your computer?
23.6.0 arm64 arm
What issue are you seeing?
What happened?
Codex Desktop sometimes starts a thread without exposing configured MCP tools to the model, even though the same MCP server is correctly detected by the Codex CLI.
In this state, the model cannot call the MCP tools because they are absent from session_meta.dynamic_tools. This looks like a Desktop tool injection / MCP manager state issue rather than an MCP server outage.
Environment
- Codex Desktop / CLI:
0.133.0-alpha.1 - OS: macOS arm64
- MCP transport:
streamable_http - Workspace has project-level
.codex/config.toml
Project config:
[mcp_servers.cysj-ai-support]
url = "https://example.com/mcp/project/mcp"
bearer_token_env_var = "AI_SUPPORT_TOKEN"
### What steps can reproduce the bug?
1. Configure a project-level streamable HTTP MCP server in `.codex/config.toml`.
2. Open the workspace in Codex Desktop.
3. Create a new thread.
4. Compare `codex mcp list` / `codex doctor` with the thread's available `dynamic_tools`.
5. Sometimes the CLI reports the MCP server as enabled, but Desktop does not inject any `mcp__...` tools into the thread.
### What is the expected behavior?
_No response_
### Additional information
_No response_This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗