agentgate
agentgate guides/use-chatgpt-codex-with-third-party-models

New in 1.4.12 · ChatGPT+Codex

Keep ChatGPT+Codex. Run DeepSeek, MiMo, and friends with real tools.

OpenAI merged ChatGPT and Codex into one desktop app. The gpt-5.6 protocol hid tool definitions inside additional_tools. Without a local hoist, third-party models never see tools—and invent fake <tool_call> text. MuxLayer 1.4.12 fixes that on your machine, not in the cloud.

one sentence

Point the merged desktop client at MuxLayer once; switch upstream providers, keep plugins when you need them, and get real tool calling on DeepSeek / MiMo / Kimi / GLM / OpenRouter and more.

What broke after the merge

The new desktop client still speaks the OpenAI Responses path that Codex users know—but tool definitions no longer always sit on the top-level tools field. They arrive as additional_tools items inside the input array.

Merged ChatGPT+Codex (gpt-5.6)
  -> Responses request
  -> tools live in input[].type = "additional_tools"   ← easy to miss
  -> naive proxy forwards empty tools
  -> DeepSeek / MiMo / … reply with fake <tool_call> text

That is not a model quality issue. The model never received the tool schema. Users see “tools broken” after updating the client even when their API keys are fine.

What MuxLayer 1.4.12 does

Problem MuxLayer fix
Tools only in additional_tools Hoist into tools before Chat / Anthropic / Gemini conversion
gpt-5.6 model names unknown Catalog + pricing for gpt-5.6, -sol, -terra, -luna; Codex mappings included
Fake <commentary> / <context_addition> tags Strip tags across SSE chunk boundaries, keep the readable text
Merged app process renamed Client restart helper recognizes main process “ChatGPT” and still falls back to Codex.app

The local path

ChatGPT+Codex desktop
  -> OpenAI-style provider entry (plugins / account path kept when you need it)
  -> http://127.0.0.1:9090  (MuxLayer)
  -> DeepSeek / Xiaomi MiMo / OpenAI / Kimi / GLM / DashScope / OpenRouter / …

MuxLayer is not a hosted reseller. Keys stay on your machine. Logs show route decisions, converted payloads, tokens, cost, and failover—so when something fails, you see where, not a black box.

Five-minute setup

  1. Install MuxLayer 1.4.12 (macOS / Windows / Linux).
  2. Add a provider (DeepSeek, MiMo, …) and paste your API key.
  3. Start the gateway — default http://127.0.0.1:9090.
  4. On Clients, click Apply Config for Codex / ChatGPT+Codex.
  5. Send a message that should call a tool. In Logs, confirm the upstream and that tools were present—not only plain text.

Need plugins and official account features? Keep the desktop app signed in; MuxLayer preserves the OpenAI-authenticated provider path while moving the model base URL to localhost. Details: Codex Desktop + plugins.

Who this is for

  • You upgraded to the merged ChatGPT+Codex app and third-party models “lost” tools overnight.
  • You want one local endpoint for Codex / Claude Code / Gemini CLI / OpenCode, not a different proxy config per app.
  • You care about failover, cost traces, and one-click switch back to official config.

FAQ

Why did tools stop working after the ChatGPT+Codex merge?

Tool definitions moved into additional_tools inside input. Proxies that only read top-level tools send an empty schema upstream.

Is MuxLayer a cloud proxy?

No. It runs locally. You bring your own provider keys. Prefer server-side multi-tenant billing? Look at LiteLLM / new-api instead—see MuxLayer vs LiteLLM.

Which version do I need?

1.4.12 or later for additional_tools hoisting and the gpt-5.6 catalog. Older builds still work as a general gateway but will not fix this protocol shape.

Related