Use CoralBricks with Codex CLI

Point OpenAI's Codex CLI at CoralBricks open models — GLM and Kimi with up to 1M context — with one provider block in config.toml.

Codex CLI is OpenAI's open-source terminal coding agent. It lets you register a custom model provider, so you can run Codex on GLM 5.2 or Kimi K3 through CoralBricks — billed to your CoralBricks account instead of an OpenAI key.

Codex speaks the Responses API. As of early 2026 Codex custom providers must use wire_api = "responses" (the older "chat" value was removed). CoralBricks serves the OpenAI Responses API natively at https://inference.coralbricks.ai/v1/responses, so no translating proxy is needed — the config below works as-is.

1. Get an API key

Create a key from your API keys page (format cb_...).

2. Add CoralBricks as a provider

Add a provider block to ~/.codex/config.toml and select it as the default:

model = "glm-5.2-fp4"
model_provider = "coralbricks"

[model_providers.coralbricks]
name = "CoralBricks"
base_url = "https://inference.coralbricks.ai/v1"
env_key = "CORALBRICKS_API_KEY"
wire_api = "responses"

env_key names the environment variable Codex reads the key from. Export your cb_... key under that name:

export CORALBRICKS_API_KEY="cb_your_key_here"

Add the line to your shell profile (~/.zshrc, ~/.bashrc) so it persists.

3. Run

codex

Codex now runs its agent loop — planning, editing files, and calling tools — on glm-5.2-fp4 through CoralBricks. Switch models per run without editing the config:

codex -c model=kimi-k3

Available models

Model	Model ID	Context	Input $/M	Output $/M
GLM 5.2	`glm-5.2-fp4`	1M	$1.12	$4.40
Kimi K3	`kimi-k3`	1M	$3.00	$15.00
GPT-OSS 120B	`gpt-oss-120b`	128K	$0.12	$0.60

Cached input tokens are always free.

Notes

Why the Responses API matters: Codex requires a provider that implements /v1/responses, not just /v1/chat/completions. Many OpenAI-compatible endpoints only expose Chat Completions and need a gateway like LiteLLM in between; CoralBricks implements Responses directly, so you skip that step.
Reasoning shows up in Codex. GLM 5.2 and Kimi K3 emit reasoning, which Codex renders inline — useful for watching an agent plan a multi-step change.
Long context is the point: GLM 5.2 and Kimi K3 accept up to 1M tokens, so large-repo Codex sessions fit without truncation.
The same key works in OpenCode, Cline, Cursor, GitHub Copilot, and anything else that speaks OpenAI-compatible providers. See the API reference for the full surface.