Codex CLI

Codex CLI is what I tell ChatGPT subscribers to try first, because it is already paid for, and an agent you do not have to justify on a card statement is the one you will actually use for a month. The source is open, which is rare for a lab's own agent. The limit is the default: OpenAI unless you write a config, since local models arrive through --oss and other endpoints through model_providers.

Pick it if you are on ChatGPT and live in a terminal. Pick Aider to choose the model, and keep this one installed for when the answer is GPT anyway.

Codex CLI sandboxes commands, sandbox_mode running from read-only to danger-full-access, but the mode is chosen once and forgotten, so the permissive setting a user picks on day three lets a bad command run on the machine. The API-key path bills per token, which lets a loop run until you notice; the Plus path rate-limits instead, which starves a long task but never surprises the card. Two failure modes, one for each way to pay.

Keep the restrictive level in a repo that matters. What it does right: codex exec is a real headless mode, so the same agent can be bounded by a script and a timeout in CI.

Codex CLI's design is in a public repository, which permits inspection of the loop rather than trust in it, and inspection shows a conventional agent: read, edit, run, repeat. A project instruction file makes conventions an explicit input rather than an inferred one, which removes one source of variance between runs. The backbone is one vendor's models only, so the scaffold cannot be separated from that vendor's model changes, and a regression in either is indistinguishable from outside.

No benchmark is published for the CLI as a scaffold. The scaffold is reproducible; the results are not yet, and the observation is that a lab could publish them tomorrow.

Codex CLI is OpenAI's answer to where the developer money goes. It is included with Plus at $20 and Pro from $100, which is bundling: the CLI raises retention rather than carrying its own price. Open source with 121,000 stars is a distribution play; the model is the business, and the CLI is the free sample that makes the model the habit.

No acquirer applies; the pivot risk is the CLI being folded into a broader app. Position: long, as a feature of a subscription you already evaluate, not as a product you bet on.

The demo is a terminal that fixes a test and explains itself. Procurement rides on the ChatGPT contract: Business is $20 per user monthly with SAML SSO, so sixty seats is $14,400 a year from a vendor procurement already knows; audit logs and SCIM are on Enterprise at a custom price, which is where a security questionnaire sends us anyway. If the org is on ChatGPT already, this is a line item, not a project.

Onboarding is a brew install. Approved with conditions: Enterprise tier for audit logs, permission levels locked down by policy, API spend capped.

Codex CLI is a lab's own agent with an Apache-2.0 license, which still surprises me. I can read the loop and the prompts, and I can fork it, which means a bug is a patch and a missing feature is a branch. MCP servers are TOML in config.toml, and AGENTS.md is a plain-text contract per repo that I check in. Then the surprise: --oss runs it against Ollama or LM Studio, and model_providers takes any OpenAI-compatible base_url, so my box is invited after all.

The grudge is that the tuning is for their models, so a local one runs worse than it deserves. Half owned, more than most.