Skip to main content

Using Codex with the RCD LLM Service

Codex can be configured to use the RCD LLM Service through a custom provider in ~/.codex/config.toml.

tip

If you are using Codex on Palmetto, see Codex on Palmetto for compute-session requirements, authentication options, and Palmetto-managed defaults.

info

Codex should use the base URL https://llm.rcd.clemson.edu/v1. The chat-completions compatibility endpoint is useful for some other clients, but Codex should be configured with the base URL.

warning

Codex depends on the service's POST /v1/responses compatibility layer. The RCD LLM Service currently provides this through a shim, so Codex support is functional but may change more often than the chat-completions endpoint.

Codex undergoes rapid development and may break things for third-party LLM providers like the RCD LLM service. If things do not work, you may have to revert to an earlier tested version. Please go ahead and report any issues you experience in as much detail as you can, and we will try our best to update the shim.

Last tested version: Codex 0.141.0.

1. Create An API Key​

First, request an allocation, sign in to the service, and create an API key:

2. Export The API Key​

Store the key in an environment variable before starting Codex:

export RCD_LLM_API_KEY="your-api-key-here"

If you want this to be available in new shells automatically, add the same line to your shell startup file such as ~/.zshrc.

3. Choose A Model​

Check the Available Models page for the exact model name to use.

4. Create a new RCD LLM Service Profile​

Create a new profile file at ~/.codex/rcd-llms.config.toml

Then add the following:

model = "<model-name>"
model_provider = "rcd-llm"
web_search = "disabled"

[feature]
image_generation = false

[model_providers.rcd-llm]
name = "RCD LLM Service"
base_url = "https://llm.rcd.clemson.edu/v1"
stream_idle_timeout_ms = 600000

[model_providers.rcd-llm.auth]
command = "bash"
args = ["-c", "echo $RCD_LLM_API_KEY"]

Replace <model-name> with an exact model name from the Models page.

info

The RCD LLM Service does not support server-side tools like web_search. They must be disabled for things to function correctly.

tip

An update to profiles was introduced recently. Make sure you are on a recent version of Codex.

Model Catalog​

Using the "command auth" configuration above should mean that Codex will automatically fetch available models from our Codex compatible models endpoint.

Alternatively, you can manually download the catalog and save it as ~/.codex/rcd_llms_model_catalog.json. Then update model_catalog_json in ~/.codex/rcd-llms.config.toml:

model_catalog_json = "~/.codex/rcd_llms_model_catalog.json"

5. Start A New Codex Session​

After saving ~/.codex/rcd-llms.config.toml, start a new Codex session with the rcd-llms profile:

codex -p rcd-llms

Notes​

  • You must still be on the Clemson network or connected through VPN.
  • If the service model list changes, update the model value in ~/.codex/config.toml.
  • Keep the API key out of Git repositories, shared shell scripts, and support tickets.