Codex

Codex CLI is OpenAI’s terminal agent. Today it cannot talk to Helmcode directly, and the reason is worth reading before you spend an afternoon on the configuration file.

Warning: this is not a problem with your key. Recent Codex versions only speak OpenAI’s Responses API — wire_api = "chat" was removed, and the binary now accepts responses and responses_websocket and nothing else. Helmcode serves /v1/chat/completions, and its /v1/responses does not yet emit the streaming events Codex waits for. Checked against codex-cli 0.154.0.

What you get if you try each of them:

What you set	What happens
`wire_api = "chat"`	Error loading config.toml: `wire_api = "chat"` is no longer supported
`wire_api = "responses"`	`stream disconnected before completion: stream closed before response.completed`

What works today — delegate to OpenCode

The same route the Claude Code page takes, and for the same reason: Codex runs commands in your terminal, and OpenCode works headless against Helmcode.

Configure OpenCode once, then ask Codex for it in plain words:

Use `opencode run --agent plan -m helmcode/glm5.3-flash` to review
src/parser.ts and tell me which cases it is not covering.

Codex runs the command, reads the answer and carries on. --agent plan keeps OpenCode read-only; without it, it starts with an agent that can edit files and run commands.

Your Codex subscription pays for the session; the model work comes out of your Helmcode key.

The configuration, for when it works

Nothing below is wrong — it is simply waiting on /v1/responses. When that endpoint streams, this is all Codex needs in ~/.codex/config.toml:

model_provider = "helmcode"
model = "glm5.3-flash"

[model_providers.helmcode]
name = "Helmcode"
base_url = "https://api.helmcode.com/v1"
env_key = "HELMCODE_API_KEY"
wire_api = "responses"

With your key in the environment, never in the file:

export HELMCODE_API_KEY="sk-your-key-here"

Two details that will still matter then:

The provider identifier cannot be openai, ollama or lmstudio, which are reserved. That is why it is called helmcode.
base_url ends at /v1 and nothing more. Do not add the endpoint path.

Known issues

/v1/responses serves fewer models than /v1/chat/completions. Even once streaming works, check the model you want is one of them.
qwen3.6 on /v1/responses rejects the request with System message must be at the beginning, which is how Codex lays out its instructions.
Everything else in this section connects directly. If you want a terminal agent on Helmcode today, OpenCode and Pi both do.