Codex-Workflows
Workflow-as-code orchestration for Codex
A plugin that orchestrates multi-agent workflows for the Codex AI coding agent with a live terminal dashboard.
Codex Workflows is a TypeScript plugin that adds a workflow orchestration layer to Codex, an AI coding agent. It connects Codex's subagents, skills, and MCP server support into a system that can run multi-step, multi-agent jobs and display their progress in a live terminal dashboard.
The plugin ships four ready-to-use workflow templates: a bounded codebase bug sweep, a deeper variant with more parallel agents, a release-diff review that checks for blockers before shipping, and a security review covering authentication, injections, secrets, and similar concerns. Workflows can run up to 64 concurrent workers and up to 2,000 total agents. State is stored on disk, separate from the target repository by default, so a run can survive a restart.
Once installed, you ask Codex to run a workflow in plain English and it calls the bundled MCP server. The server starts the job and opens a dashboard in your terminal showing phases, individual agent rows, token and tool usage, recent activity, and the final report location. From the terminal you can pause, resume, stop, restart specific agents, or save a completed run under a name you choose.
Model selection can be set globally or per phase. You can point the bug-finding phases at a cheaper model and the final synthesis step at a stronger one, for example, using a single parameter. Model names are checked against the local Codex catalog before any agents start, so a typo produces an error with a suggestion rather than a silent wrong-model run.
Workflow scripts define phases and agent prompts, and they run inside a QuickJS sandbox that does not give them direct filesystem, shell, or network access. The plugin is committed to the repository in a built form, so users can install it from GitHub without needing a local build step. It is released under the MIT license.
Where it fits
- Run a bounded bug sweep across a codebase using multiple parallel agents
- Check a release diff for blockers before shipping
- Run a security review covering authentication, injections, and secrets
- Watch a live terminal dashboard showing agent progress, tokens, and activity