AI Potluck
Product / UX / Orchestration & agents

Codex CLI

OpenAI

Terminal coding agent from OpenAI, written in Rust. It reads a codebase, proposes a plan and executes multi-file changes with sandboxed command execution, and documents writable roots, permission profiles, agent approvals, auto-review, MCP connectors and skills. Desktop builds are offered for Linux and Windows.

Codex runs against either a ChatGPT plan or an API key. Verified 2026-08-13 via the OpenAI Codex CLI documentation.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI, the CLI agent)
source
public(GitHub)
note
agent is OSS but optimized for/defaults to OpenAI models (provider-leaning, not license-restricted)
core-gated
ungated(one Apache-2.0 LICENSE at the root governs the whole tree

Apache-2.0, an OSI-approved license, with public Rust source (96.1% Rust); OpenAI's own docs call it open source. The backing models are proprietary, but the orchestration agent scored here is fully open. A full recursive tree of openai/codex at main, untruncated, carries exactly one governing LICENSE - the root Apache-2.0 - and the only other license files cover vendored third-party dependencies (third_party/wezterm, codex-rs/vendor/bubblewrap) and sample skill assets. There is no enterprise or commercial directory, which is the test for whether a core has been carved out. What you pay for is model access: the README directs you to sign in with a ChatGPT Plus, Pro, Business, Edu or Enterprise plan, or with an API key. That is a hosted service sold alongside the published agent rather than a feature withheld from it, so nothing gates the core.

Adoption

4 high confidence
4.0

OpenAI has disclosed that Codex passed 4 million weekly developers, and the Wikipedia article on the agent cites more than 6 million weekly active users - both firmly in the 1-10M band on disclosed developer count. The OpenAI announcement carrying the original figure is no longer fetchable, since openai.com answers 403 to a non-browser request, so the band rests on the larger Wikipedia-cited number instead. What is banded is a weekly active count rather than the monthly count the scale is written in, so if anything it understates a monthly reading.

Capability

4 medium confidence
4.0

Capable, heavily-used coding agent backed by frontier OpenAI models, but no standalone agent SWE-bench placement published. Scored on feature matrix; high but not the open-agent benchmark frontier-definer (OpenHands 68.4%).

  • https://developers.openai.com/codex/cli recorded 2026-08-13

    autonomous terminal agent documented across sandboxing and writable roots, permissions and profiles, agent approvals and security, auto-review, MCP and connectors, skills and plugins, and Linux/Windows desktop apps; no benchmark figure published for the agent

Verified 2026-08-11