MCP Errors: 0.12.0 was last version to load and reliably run most _reference_ MCP servers
What version of Codex is running?
0.13.0 with errors being emitted in 0.14.0+
Which model were you using?
gpt-oss
What platform is your computer?
Linux x86_64
What steps can reproduce the bug?
any versions work reliably? errors?
0.12.0 yes, [no errors]
0.13.0 no, [no errors]
0.14.0 no, Failed to convert MCP tool to OpenAI tool: Error....
0.15.0 no, Failed to convert MCP tool to OpenAI tool: Error....
0.16.0 no, Failed to convert MCP tool to OpenAI tool: Error....
0.17.0 missing release
0.18.0 missing release
0.19.0 no, Failed to convert MCP tool to OpenAI tool: Error....
0.20.0 no, Failed to convert MCP tool to OpenAI tool: Error....
What is the expected behavior?
when using the tui, successful tool calling is normally noted in the log without a reload error on all MCP servers for each prompt
What do you see instead?
errors on most of the basic set of so-called "Reference" MCP Servers, mcp/git, for example:
_Failed to convert "git__git_branch" MCP tool to OpenAI tool: Error("missing field type", line: 0, column: 0)_
Additional information
_No response_
3 Comments
"Reference" MPC Servers being https://github.com/modelcontextprotocol/servers?tab=readme-ov-file#-reference-servers ...such as "git", "sequential-thinking", and "fetch". But also seeing this with many of the popular Docker Hub -curated MCP servers under the "mcp/" registry namespace.
confirmed PR 1975 ( https://github.com/openai/codex/pull/1975 ) allows the tools to load
this might be a duplicate of #1975 so, closing.
...ran a local cargo build of the rust binary in x86_64 linux after that was merged to main, and seeing _much_ useful behavior - especially outside the TUI with one-off "codex exec" runs ( critical to shell scripted workflows). feels much more solid.
looking forward to the next release!