Run Codex CLI with open-weight models

The Codex CLI runs with open-weight models on Munito AI: add Munito AI as a model provider, and Codex calls the Munito AI Responses API. Create an API key. Export the key as MUNITO_API_KEY:

export MUNITO_API_KEY=munitoai_...

Add this configuration to ~/.codex/config.toml:

model = "zai-org/glm-5.3-flash"
model_provider = "munito"

[model_providers.munito]
name = "Munito"
base_url = "https://api.munito.ai/inference/v1"
env_key = "MUNITO_API_KEY"
wire_api = "responses"

Run codex. The wire_api = "responses" setting connects Codex to the Munito AI Responses API. The command codex exec "<task>" runs one task without a terminal UI. All stock Codex tools work: shell commands, file edits and sub-agents. Munito AI runs no web search, so the model answers without a search.

Use a regional host to keep processing in the EU or the US. Replace the base_url with the URL for that region:

• EU: https://api.eu.munito.ai/inference/v1
• US: https://api.us.munito.ai/inference/v1

Change model to use another model. See the models that your key can use.

Model metadataCopy Anchor Link

Codex has no metadata for a Munito AI model. It prints this warning:

Model metadata for `zai-org/glm-5.3-flash` not found

With the fallback, Codex does not know the context window of the model, but the run still works. The model_context_window setting alone does not remove the warning. Use model_catalog_json to supply the metadata and remove the warning.

The command needs jq. It copies one entry of the catalog that ships with Codex and changes the fields for Munito AI. The catalog file replaces the built-in list. Run this command before you add model_catalog_json:

codex debug models | jq '{models: [.models[] | select(.slug == "gpt-5.5")
  | .slug = "zai-org/glm-5.3-flash"
  | .display_name = "GLM-5.3 Flash"
  | .description = "GLM-5.3 Flash on Munito"
  | .context_window = 1048576
  | .max_context_window = 1048576
  | .supports_search_tool = false
  | .support_verbosity = false
  | .upgrade = null
  | .service_tiers = []
  | .additional_speed_tiers = []
  | del(.apply_patch_tool_type, .default_verbosity)]}' > ~/.codex/munito-models.json

GLM-5.3 Flash has a context window of 1,048,576 tokens. The other changes disable request features that Munito AI does not accept: the tool search tool, the freeform patch tool and the verbosity setting.

Add this top-level key to config.toml, above [model_providers.munito]:

model_catalog_json = "/Users/you/.codex/munito-models.json"

Use the absolute path of your home directory in place of /Users/you. The catalog then lists only the models in that file.

Reasoning effortCopy Anchor Link

Set model_reasoning_effort = "high" at the top of config.toml, above [model_providers.munito]. Use codex -c model_reasoning_effort="high" for a command-line override. Codex sends minimal, low, medium, high or xhigh. Each model declares its own levels. Munito AI moves any other level to the nearest declared level, and to the lower level on a tie.

Codex cannot send max.