Codex Unleashed

Codex Unleashed is a public fork of OpenAI Codex CLI kept close to the rust-v0.87.x CLI lineage while remaining usable with current OpenAI models.

It installs as codex-unleashed, not codex, and stores its own user state in ~/.codex-unleashed by default. This lets it live next to the regular Codex CLI without sharing auth, config, logs, sessions, caches, or helper files.

Why this fork exists

This fork preserves a style of Codex use that became harder to keep stable as later upstream versions moved toward faster iteration, more automatic orchestration, and broader multi-agent workflows.

Those directions are useful, but they change the operating model. Deep agentic coding often depends on one long-running agent maintaining judgment, context, and accountability over a task. Splitting work across agents is not free: shared context has to be selected or summarized, ownership has to stay visible, and orchestration has to avoid silently changing model, reasoning effort, or task state. When those boundaries are unclear, the result can feel less agentic even if the system is faster.

The degradation this fork is designed around is not simply "newer is worse" or "multiple agents are bad." It is the practical loss of continuity that shows up when a coding assistant optimizes for dispatch and turnaround before context ownership and orchestration semantics are mature enough for deep work.

Codex Unleashed takes the opposite default stance: depth over speed, explicit configuration over hidden routing, and predictable long-running behavior over automatic handoff. The public branch is intentionally conservative. It keeps the 0.87-era CLI recognizable, isolates the fork from upstream codex, suppresses upstream update prompts, and updates model discovery so current OpenAI models remain available.

The fork also keeps model cost and reasoning tradeoffs visible. You can pin a frontier model for hard work, choose a cheaper model for routine tasks, and set reasoning effort explicitly instead of treating model choice as an opaque default.

See Rationale for the longer design note.

Install

From npm, once a release has been published:

npm install -g @lpalbou/codex-unleashed

macOS/Linux:

curl -fsSL https://raw.githubusercontent.com/lpalbou/codex-unleashed/main/scripts/install/codex-unleashed.sh | sh

Windows PowerShell:

irm https://raw.githubusercontent.com/lpalbou/codex-unleashed/main/scripts/install/codex-unleashed.ps1 | iex

The installer builds from source, so it requires git, cargo, and a Rust toolchain. To install from this checkout instead:

cd codex-rs
cargo install --locked --path cli --bin codex-unleashed

Run

codex-unleashed
codex-unleashed login
codex-unleashed --model gpt-5.5

Global configuration lives at ~/.codex-unleashed/config.toml. Override the home directory with CODEX_UNLEASHED_HOME=/path/to/home.

The bundled fallback model catalog prefers gpt-5.5, then gpt-5.4, then gpt-5.4-mini. Local/source builds use a compatibility client-version floor for OpenAI /models discovery so the fork can continue to see current model catalog entries.

This public branch does not publish private experimental branches. Private work stays outside this repository; public releases are cut from the public main branch only.

Docs

• Rationale
• Installing and building
• Documentation index
• npm publishing
• Configuration
• Authentication
• Sandboxing
• Contributing

This repository is licensed under the Apache-2.0 License.