Codex CLI
Run OpenAI Codex as a role-capable agent inside a managed pane, with backend-aware readiness and active-work detection.
$ codex
Turn a task into a visible, reviewable pipeline. AgentFlow coordinates specialized coding agents across planning, implementation, review, and verification—using Codex CLI, Claude Code CLI, or Cursor CLI while you stay in control.
AgentFlow is a macOS desktop app that coordinates Codex CLI, Claude Code CLI, and Cursor CLI through tmux. Install tmux and at least one supported CLI, then open your repository as a workspace.
Download for macOS DMG · Apple siliconAgentFlow uses tmux to host and observe agent panes.
brew install tmux
Install and sign in to Codex CLI, Claude Code CLI, or Cursor CLI. AgentFlow needs at least one available backend.
$ codex
$ claude
$ agent
Open the downloaded DMG and drag AgentFlow into Applications.
Launch AgentFlow, choose your repository, map agent backends to panes, and start the runtime.
Assign a backend to each tmux pane and let AgentFlow schedule work by role. Use one CLI everywhere, or mix backends across your agent grid.
Run OpenAI Codex as a role-capable agent inside a managed pane, with backend-aware readiness and active-work detection.
$ codex
Use Claude Code for planning, implementation, or review roles while AgentFlow handles dispatch and pane lifecycle.
$ claude
Run Cursor Agent in autonomous mode with first-run, readiness, and usage-state awareness built into the control plane.
$ agent
Install and authenticate at least one supported CLI before starting AgentFlow. Every configured pane must reference an available backend.
AgentFlow gives multi-agent development a durable workflow, a live operating surface, and explicit moments for human judgment.
Select planning, breakdown, and implementation stages. AgentFlow inserts the right review jobs, optional human checks, and final verification gates around them.
A pane-centric tmux grid exposes readiness, active work, and live terminal snapshots without losing the task-level narrative.
A durable orchestrator routes jobs, reviews, retries, and recovery while preserving the context and artifacts of every attempt.
Inspect plans, request changes, verify the result, pause a task, or recover a failed run. Automation remains legible and reversible.
Give AgentFlow the outcome. It turns that intent into a coordinated sequence of specialist work, review, and human verification.
Install AgentFlowChoose a project, describe the outcome, attach context, and select the stages and checks this change needs.
AgentFlow dispatches compatible panes, advances the pipeline, and pairs work with role-specific reviews.
Inspect the full DAG and artifacts, handle exceptions, then pass the human gate when the change is ready.