Codex Workflows MCP Server
Enables durable multi-agent workflow runs for Codex with a live terminal dashboard, restartable agents, and bug-finding templates, all controllable through natural language.
README
Codex Workflows
Claude Code has workflows. Codex has subagents, skills, plugins, and MCP.
codex-workflows connects those pieces into a workflow-as-code runtime for
Codex: durable multi-agent runs, a live terminal dashboard, restartable agents,
saved workflow scripts, and bug-finding templates that are meant to be used on
real repos.
This is not a native Codex /workflows command. Codex does not expose that
plugin API today. This project ships the closest available surface: a Codex
plugin with an MCP server and a cwf terminal UI.
Install
Once this repo is public, install it as a Codex marketplace:
codex plugin marketplace add robzilla1738/Codex-Workflows
codex plugin add codex-workflows@codex-workflows
Start a new Codex thread after installing. Then ask:
Use codex-workflows to run the bug-sweep workflow with adapter auto.
Open the live dashboard automatically and tell me the run id and status.
Codex will call the bundled MCP server. The server starts the workflow, opens the live TUI in your default terminal, and returns a run id you can inspect from Codex or from the CLI.
What it includes
bug-sweep: bounded codebase review, adversarial verification, repro planning, and synthesis.bug-sweep-deep: larger opt-in fanout for deeper bug hunts.release-diff-review: release-blocker review against a branch or diff.security-auth-review: auth, permission, injection, secret, sandbox, and MCP boundary review.- Live dashboard with phases, agent rows, token/tool/time metrics, recent worker activity, detail view, pause/resume/stop/restart/save controls, and final report path. Token totals show as pending until Codex emits usage metadata; command/tool/message/stderr activity updates as events arrive.
- Workflow definitions can explicitly request up to 64 concurrent workers and 2000 total agents.
- Durable run state under
${CODEX_HOME:-~/.codex}/codex-workflows/projects/<project-hash>/runs/<run-id>/by default, so read-only bug hunts do not dirty the target repo. - Project-local storage remains available with
storageScope: "project"or--storage-scope project. - Downstream verify, probe, and synthesize agents receive prior phase findings as explicit context instead of guessing at earlier results.
- Detached worker heartbeats let status reads and dashboards identify orphaned runs when a background workflow process dies.
- Isolated workflow script loading through QuickJS. Workflow scripts define phases and agent prompts; they do not get direct filesystem, shell, network, process, or Node built-in access.
Model controls
You can route models globally or by phase/agent:
Use codex-workflows to run bug-sweep with adapter auto,
model gpt-5.5, reasoning xhigh,
and modelMap {"find":"gpt-5.4-mini","synthesize":"gpt-5.5"}.
Caller overrides win over workflow defaults:
model: default worker model.reasoning: default Codex reasoning effort.modelMap: phase id, agent id, orphase:agentoverrides.promptSuffix: extra instruction appended to every worker.
Model names are validated against the local codex debug models catalog before
any agents launch. Short names such as 5.4-mini are rejected with a suggestion
like gpt-5.4-mini; they are not silently rewritten.
Bug-finding workflows default to read-only workers unless a workflow explicitly declares otherwise.
Local development
pnpm install
pnpm validate:plugin
pnpm test
pnpm cwf validate workflows/bug-sweep.workflow.js
pnpm cwf run workflows/bug-sweep.workflow.js --watch --adapter simulate
pnpm cwf run workflows/bug-sweep.workflow.js --watch --adapter auto --model gpt-5.4-mini
pnpm cwf run workflows/bug-sweep-deep.workflow.js --watch --adapter auto --model gpt-5.4-mini
Useful local commands:
pnpm cwf workflows
pnpm cwf watch <run-id>
pnpm cwf pause <run-id>
pnpm cwf resume <run-id>
pnpm cwf stop <run-id>
pnpm cwf stop-agent <run-id> <agent-id>
pnpm cwf restart-agent <run-id> <agent-id>
pnpm cwf save <run-id> --name release-diff-review-v2
The generated plugin lives at plugins/codex-workflows. Do not edit the bundled
dist files by hand. Change the TypeScript packages, then run:
pnpm validate:plugin
That builds the packages, regenerates the plugin bundle, and checks that no machine-local paths leaked into the distributable plugin.
Plugin layout
.agents/plugins/marketplace.json
plugins/codex-workflows/
.codex-plugin/plugin.json
.mcp.json
dist/
skills/codex-workflows/SKILL.md
workflows/*.workflow.js
The plugin bundle is committed so users can install from GitHub without running
pnpm install.
Docs
Support
If this saves you time, you can support the work here: buymeacoffee.com/robcourson.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.