MCP unsupported call

Open 💬 2 comments Opened Jul 2, 2026 by sunwei-yhc

What version of Codex CLI is running?

codex-cli 0.142.5

What subscription do you have?

No

Which model were you using?

MinimaxM3

What platform is your computer?

macOS-M3Max-版本:26.5.1

What terminal emulator and version are you using (if applicable)?

Ghostty

Codex doctor report

three MCP services, codegraph, context7, and chrome_devtools.
All three calls return 'unsupported call'

What issue are you seeing?

After connecting to the Minimax M3 model, the model can be used normally, but the MCP tool in codexcli cannot be used properly, with an error message: parallel testing of the three MCP services, codegraph, context7, and chrome_devtools.
All three calls return 'unsupported call'

What steps can reproduce the bug?

<img width="3262" height="1854" alt="Image" src="https://github.com/user-attachments/assets/ce8723df-9c49-4955-be41-0c363fccf4b0" />

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

2 Comments

tangweigang-jpg · 1 month ago

The strongest signal here is that three unrelated MCP servers all fail with the same unsupported call only in the custom-model path. I would try to separate MCP server setup from model/tool-call compatibility before debugging codegraph/context7/chrome_devtools individually.

A small repro that would help maintainers: run the same Codex config once with the default OpenAI model and once with MinimaxM3, then capture the doctor output plus the exact provider response around the first attempted tool call. If the default model can invoke one trivial MCP tool but MinimaxM3 cannot, this is probably in the custom provider's tool-call adapter/shape rather than in the MCP servers themselves.

crazyi · 1 month ago

Same root cause, with a more specific reproduction that confirms the Apps SDK angle.

Filed #31354 with full RCA: all 7 unsupported call hits in ~/.codex/logs_2.sqlite come from MiniMax-M3 (same model as you), and the failure happens at core/src/tools/router.rs:185 (dispatch_tool_call_with_terminal_outcome) before the MCP child process is even reached — verified by running the MCP probe script standalone and getting OK: ....

The interesting twist that distinguishes this from a generic MCP failure: even non-widget MCP tools on the same Apps SDK MCP server also fail (e.g. get_cowart_canvas_state, save_cowart_canvas_state, ...), because Codex namespaces the whole server as an Apps SDK server based on _meta.ui.resourceUri / openai/outputTemplate exposed by any of its tools. So this equally affects Cowart, Canva, Computer Use, and any future Apps-style MCP server used with a custom Responses provider.

A clean repro (Cowart v0.1.11 + Codex CLI 0.142.5 + MiniMax-M3) and the exact dispatcher path are in #31354. Adding cowart@personal to the three MCPs in your doctor output should reproduce this 1:1.