Model selection

The MCP server normally omits --model, so Codex CLI selects the model from your account and $CODEX_HOME/config.toml. Pass model only when a task needs a deliberate override.

This list was verified against Codex CLI 0.159.2 and its model cache on 2026-09-30, with a live one-token run per slug. Availability still depends on account, workspace, and rollout.

Recommended models

Model	Best for	Guidance
gpt-6.1-sol	Everyday coding and agentic work	Latest workhorse; needs Codex CLI 0.159+
gpt-6-astra	Complex, demanding, high-value work	Frontier intelligence; use when it matters
gpt-6-sol	Everyday work on an older CLI	Previous-generation workhorse
gpt-6-luna	Clear, repeatable, high-volume work	Fast and affordable
gpt-5.6-sol	Older-generation work	Still available
gpt-5.6-terra	Straightforward work	Older balanced model
gpt-5.6-luna	Simple, well-scoped tasks	Older fast, efficient model
gpt-5.5	Legacy compatibility	Codex steers callers to gpt-5.6-sol

Reach for Astra when the problem is hard or open-ended, use Sol as the everyday workhorse, and use Luna when the task is specific and success is easy to verify.

gpt-6.1-sol is rejected by Codex CLI 0.157.x with "not supported when using Codex with a ChatGPT account" — the same 400 an unknown slug gets. Upgrade with npm install -g @openai/codex@latest. gpt-5.4-mini was removed upstream.

Pass a concrete slug. The bare moving aliases gpt-6 and gpt-5.6 are not accepted — the API rejects them with an HTTP 400.

Reasoning effort

Supported MCP values are low, medium, high, xhigh, max, and ultra. The selected model and account determine which values are actually available.

• Start with medium.
• Use low for quick, tightly scoped tasks.
• Use high or xhigh for difficult multi-step work.
• Use max only for the hardest single-agent problems where latency matters less.
• Use ultra only when the task can benefit from automatic delegation to subagents.

Most tasks do not need max or ultra. In the verified CLI cache, Sol and Terra expose both; Luna exposes max but not ultra.

Examples

Direct CLI:

codex exec -m gpt-6-astra "Review the current changes"
codex exec -m gpt-5.6-terra "Refactor the request parser"
codex exec -m gpt-5.6-luna "Classify these build errors"

MCP invocation:

{
  "prompt": "Audit @src for command-building regressions",
  "model": "gpt-5.6-sol",
  "reasoningEffort": "high"
}

For the default model, omit model:

{
  "prompt": "Explain @src/utils/codexCommandBuilder.ts"
}

Set a persistent default in $CODEX_HOME/config.toml:

model = "gpt-5.6-sol"
model_reasoning_effort = "medium"

Local models

Local provider model names are passed through without OpenAI model validation:

{
  "prompt": "Explain @src",
  "localProvider": "ollama",
  "model": "qwen3:8b"
}

There is no MCP-side automatic fallback chain. When no override is supplied, Codex CLI remains the source of truth. An unknown explicit model is passed through so newer rollouts and local providers are not blocked by this package.

See the official Codex model guide for current availability and recommendations.