Codex Starter Kit

A ready-to-use baseline for OpenAI Codex CLI: agents, skills, hooks, MCP, plugins, and global AGENTS.md.

Install it once, restart Codex, and get a practical working setup for development, review, QA, DevOps, design, product, data, and security work.

English · Русская версия

GitHub Repository · Telegram: @pomogay_marketing

What This Is

codex-starter-kit turns a clean Codex CLI setup into a ready working environment.

It includes:

• 62 custom Codex agents in agents/*.toml
• 101 reusable skills in skills/*/SKILL.md
• global Codex working rules in templates/AGENTS.md
• a shell safety hook for risky commands
• default command approval rules in rules/default.rules
• a baseline ~/.codex/config.toml with GitHub plus bundled local Superpowers and Claude Companion plugins
• public docs MCP servers: Context7, Vue, Nuxt UI, and Nuxt
• recommended local MCP routes for Serena, GitNexus, Postgres, Open Design, and claude-mem
• bundled VKirill/superpowers fork in plugins/superpowers
• repo-local Claude Companion plugin prototype in plugins/claude-companion for interactive Claude Code reviews through tmux
• an installer with --dry-run, backups, runtime refresh, and validation

The point is simple: you do not rebuild a Codex workflow from scratch. Install the baseline, verify it, then keep project-specific rules in each repository's local AGENTS.md.

Fast Path

If you want Codex to install this kit for you, open Codex on the target machine and paste this prompt.

Install Codex Starter Kit for OpenAI Codex CLI and verify that the installation is safe.

Repository:
https://github.com/VKirill/codex-starter-kit

Goal:
- make this starter kit the baseline Codex setup on this machine
- install global ~/.codex/AGENTS.md from templates/AGENTS.md
- install custom agents into ~/.codex/agents
- install skills into ~/.agents/skills
- install safety hooks into ~/.codex/hooks and ~/.codex/hooks.json
- install safe command approval rules into ~/.codex/rules
- install the recommended ~/.codex/config.toml from templates/config.recommended.toml
- enable GitHub plus bundled local Superpowers and Claude Companion plugin entries through config.toml
- enable public docs MCP servers for Context7, Vue, Nuxt UI, and Nuxt
- check and document recommended local MCP/plugin routes for Serena, GitNexus, Postgres, Open Design, and claude-mem
- perform full local workflow self-setup when safe: Serena, GitNexus, and claude-mem
- enable Claude Companion only when Claude Code CLI is already installed; do not install Claude or tmux automatically
- include GitHub/source links for every enabled plugin and recommended MCP/plugin route
- preserve old files with timestamped .bak-* backups

Work step by step:
1. If the repository is missing, clone it into ~/projects/codex-starter-kit.
2. If the repository already exists, enter it and check the current git status.
3. Read README.md, install.py, and templates/config.recommended.toml.
4. Run the pack validation:
   python3 scripts/validate-pack.py
5. Run a dry run:
   ./install.sh --dry-run
6. Show me which paths will be replaced, then count agents and skills separately.
7. If the dry run looks safe, run the install with backups:
   ./install.sh
8. Verify that ~/.codex/config.toml contains GitHub, superpowers@codex-starter-kit, and claude-companion@codex-starter-kit plugins, plus MCP servers context7, vue-docs, nuxt-ui-remote, and nuxt-remote.
9. Report that Serena, GitNexus, Postgres, Open Design, and claude-mem are recommended local/plugin routes for the full starter-kit workflow.
10. For each enabled plugin and recommended MCP/plugin route, show its GitHub/source link from README.md or templates/config.recommended.toml.
11. Run full local workflow preflight:
   - detect OS and package manager: apt, dnf, pacman, zypper, brew, or none
   - check commands: tmux, claude, uv, uvx, node, npm, npx, gitnexus
   - check Codex runtime: codex, `codex plugin marketplace upgrade`, `codex mcp list`
   - report missing tools before installing anything
12. If Claude Code CLI is missing, do not install Claude and do not install tmux for Claude Companion. Report that Claude Companion is disabled on this machine until the user installs and authenticates Claude Code manually.
13. If Claude Code CLI exists, verify it with `claude --version`. Then check tmux. If tmux is missing, do not install it automatically; show the manual install command for the detected package manager:
   - Ubuntu/Debian: `sudo apt-get update && sudo apt-get install -y tmux`
   - Fedora/RHEL: `sudo dnf install -y tmux`
   - Arch: `sudo pacman -S --needed tmux`
   - openSUSE: `sudo zypper install -y tmux`
   - macOS: `brew install tmux`
14. Do not read, copy, parse, print, or move Claude credentials. Use Claude Code's own authentication flow only. If Claude is installed but not authenticated, open an interactive `claude` session and ask the user to run `/login`; then re-check. Do not inspect `~/.claude/.credentials.json`, cookies, keychains, bearer tokens, or API keys.
15. If uv/uvx is missing and Serena is requested or absent, install uv after approval using the official Astral installer or the system package manager. Then install and initialize Serena:
   `uv tool install -p 3.13 serena-agent@latest --prerelease=allow`
   `serena init`
   Verify that `serena` is on PATH. Do not hardcode private ports or bearer tokens in public starter-kit files.
16. If GitNexus is missing, install or use it after approval:
   `npm install -g gitnexus`
   or for one-off MCP use:
   `npx -y gitnexus@latest mcp`
   Then run safe checks: `gitnexus --help`, `gitnexus status` in the current repo if applicable. Ask before indexing large repositories. If the user approves indexing this starter-kit repo, run `gitnexus analyze` from the repo root.
17. If claude-mem is missing and the user wants memory continuity, install it after approval:
   `npx claude-mem@latest install`
   Then restart Codex and verify available MCP/tools if safe. If install fails or appears unstable, report the failure and leave it disabled.
18. If recommended local MCP/plugin routes are already installed and safe to verify, check them with `codex mcp list` or their own read-only status command. Do not write private ports, local paths, bearer tokens, or database credentials into public starter-kit files.
19. If the codex command is available, run:
   codex plugin marketplace upgrade
   codex mcp list
20. Validate installed agents:
   ./install.sh --validate-only
21. At the end, briefly summarize what changed, where backups were written, which tools were installed, which recommended MCP/plugin routes are active or missing, what still needs manual login, and that Codex must be restarted.

Safety rules:
- do not delete ~/.codex, ~/.agents, or existing agents/skills without backups
- do not use --force or --no-backup without my explicit approval
- do not copy secrets, bearer tokens, private MCP settings, or local database credentials
- do not install Claude Code or tmux automatically; Claude Companion is opt-in for machines where Claude already exists
- do not read Claude credentials directly; always delegate authentication to Claude Code
- ask before sudo, package-manager installs, global npm installs, `curl | sh`, large repo indexing, daemon startup, or service changes
- prefer read-only checks before mutation; if an install fails, stop and show the smallest safe fix
- if a command needs elevated access or network approval, explain why first
- if a check fails, stop, show the error, and suggest the smallest safe fix

Manual Install

Clone the repository:

git clone https://github.com/VKirill/codex-starter-kit ~/projects/codex-starter-kit
cd ~/projects/codex-starter-kit

Validate the pack:

python3 scripts/validate-pack.py

Preview what the installer will replace:

./install.sh --dry-run

Install the baseline with backups:

./install.sh

Validate installed agents:

./install.sh --validate-only

Restart Codex after installation. Global instructions, agents, skills, hooks, and plugins may not load until the next Codex session.

What Gets Installed

Path	What goes there	Why
~/.codex/AGENTS.md	global working rules	consistent Codex behavior across projects
~/.codex/agents/	62 custom subagents	roles for development, review, QA, DevOps, product, design, and copywriting
~/.agents/skills/	101 skills	reusable instructions for tasks and domains
~/.codex/hooks/	safety and handoff hook scripts	classifies user prompts, guards risky shell commands, auto-approves known safe permission prompts, and nudges verification after installs/failures
~/.codex/hooks.json	hook config	connects UserPromptSubmit, PermissionRequest, PreToolUse, and PostToolUse hooks to Codex
~/.codex/rules/	command approval rules	auto-approves common read-only development, Linux, package metadata, and diagnostics commands
~/.codex/config.toml	baseline config	plugins, MCP servers, approvals, docs discovery

By default, the installer replaces managed paths only after moving existing files to timestamped .bak-* backups.

Agent Map

Agents live in agents/*.toml. Each agent has a narrow role, its own model_reasoning_effort, nickname candidates, and a focused skill set.

Category	Use it for	Agents
	handoff intake scoring, task ledgers, planning, workflows, coordination	agents_orchestrator, project_manager_senior, specialized_workflow_architect, automation_governance_architect
	backend, frontend, mobile, CMS, architecture, code	engineering_backend_architect, engineering_frontend_developer, engineering_senior_developer, engineering_software_architect, engineering_minimal_change_engineer, engineering_rapid_prototyper, engineering_mobile_app_builder, engineering_cms_developer, engineering_codebase_onboarding_engineer, engineering_code_reviewer, engineering_technical_writer, engineering_git_workflow_master, lsp_index_engineer, terminal_integration_specialist, specialized_mcp_builder, specialized_developer_advocate
	ML, data pipelines, databases, email/audio intelligence	engineering_ai_engineer, engineering_ai_data_remediation_engineer, engineering_data_engineer, engineering_database_optimizer, engineering_email_intelligence_engineer, engineering_voice_ai_integration_engineer, specialized_model_qa
	infrastructure, incidents, security, compliance	engineering_devops_automator, engineering_sre, engineering_security_engineer, engineering_threat_detection_engineer, engineering_incident_response_commander, engineering_autonomous_optimization_architect, compliance_auditor
	checks, evidence, accessibility, performance	testing_api_tester, testing_evidence_collector, testing_accessibility_auditor, testing_performance_benchmarker, testing_reality_checker, testing_tool_evaluator, testing_test_results_analyzer, testing_workflow_optimizer
	roadmaps, feedback, sprints, experiments, delivery	product_manager, product_feedback_synthesizer, product_trend_researcher, product_sprint_prioritizer, product_behavioral_nudge_engine, project_management_project_shepherd, project_management_experiment_tracker, project_management_jira_workflow_steward, project_management_studio_operations, project_management_studio_producer
	UI, UX, brand, visuals, research	design_ui_designer, design_ux_architect, design_ux_researcher, design_brand_guardian, design_visual_storyteller, design_image_prompt_engineer, design_inclusive_visuals_specialist, design_whimsy_injector
	knowledge bases, notes, cross-domain reasoning	zk_steward

Skill Groups

Skills live in skills/. The installer copies them to ~/.agents/skills, where agents can use them for sharper task behavior.

Group	Examples
Process	planning-methodology, task-decomposition, testing-patterns, bug-hunter, code-review-checklist
Frontend	frontend-developer, react-patterns, nextjs-best-practices, vue-developer, ui-designer, playwright-skill
Backend	nodejs-expert, fastify-pro, fastapi-pro, api-patterns, auth-implementation-patterns, graphql
Data	postgresql, database-design, prisma-expert, drizzle-orm-expert, redis-patterns, data-engineer
Ops	docker-expert, terraform-specialist, linux-sysadmin, github-actions-templates, server-management
Security	security-audit, backend-security-coder, find-bugs, incident-responder
Product and Docs	copywriter, ru-text, roadmap-methodology, goal-achievement-review, software-architecture

Subagents use role-based allowlist emulation through [[skills.config]] enabled = false, so each role sees a focused skill menu instead of the whole library.

Safety Model

The installer runs in baseline mode: it makes this starter kit the main Codex setup on the machine.

Default safeguards:

• ./install.sh --dry-run previews replacements without writing
• old managed paths are moved to timestamped .bak-* backups
• agent TOML files are validated after install
• skills.config paths are rewritten for your home directory
• codex plugin marketplace upgrade and codex mcp list run only when codex is available in PATH
• the local codex-starter-kit plugin marketplace is registered dynamically from the install path, and bundled plugins such as superpowers@codex-starter-kit and claude-companion@codex-starter-kit are enabled when their metadata exists
• bundled local plugins are also copied into the Codex plugin cache, so their skills are visible in new Codex sessions without a separate manual install step
• rules/default.rules reduces routine approval prompts for read-only commands such as package metadata checks, Linux inspection commands, service status checks, Docker/Kubernetes/Terraform read-only inspection, and GitHub CLI view/list commands
• The baseline uses approval_policy = "never" with workspace-write for full trusted auto-edit handoff flow; critical commands still go through rules/hooks.
• Recoverable deletion with gio trash <path> is auto-approved. Permanent deletion commands such as rm, rmdir, and unlink remain blocked.
• cp is auto-approved for autonomous handoff file-copy workflows; sandbox scope and safety hooks still constrain where it can write.
• printf, python, python3, and broad pnpm commands are auto-approved for autonomous handoff workflows; critical mutations such as package publish/deploy/release remain blocked by the safety hook.
• npm workspace forms (npm --workspace, npm -w, npm --workspaces, npm --prefix) and pnpm workspace forms (pnpm --filter, pnpm -F, pnpm --recursive, pnpm -r, pnpm --dir, pnpm -C) are approved for handoff development workflows
• hooks/handoff-intake-classifier.py classifies user prompts on UserPromptSubmit only when ~/.codex/private/handoff-classifier.env exists. With that private file present, it works offline with deterministic fallback and uses the LLM as the primary classifier when OPENAI_API_KEY and HANDOFF_CLASSIFIER_MODEL are set. The LLM path requests strict Responses API Structured Outputs (text.format JSON Schema) and normalizes typed booleans such as should_edit, requires_release_flow, requires_worktree_gate, and requires_gitnexus_impact; plain JSON parsing remains only as a compatibility fallback. The rendered hook context is compact English.
• For implementation prompts, the classifier emits a normalized engineering brief with professional architecture vocabulary and, when Codex provides a working directory, a compact repo profile from root and nearest-workspace package.json, AGENTS.md, runtime engines, common monorepo layout signals, and an allowlist of architecture-significant dependency versions such as frameworks, build tools, test tools, ORMs, databases, queues, contract libraries, UI kits, and observability libraries.
• Short follow-up prompts such as доработай, продолжай, ещё, тогда, or сделай так can use the previous local hook context before calling the LLM. State is stored in ~/.codex/private/handoff-classifier-state.json when writable, otherwise ~/.codex/memories/handoff-classifier-state.json; override with HANDOFF_CLASSIFIER_STATE_PATH. The state is capped, stored with 0600 permissions, scoped by session/repo when possible, and redacts common API keys/tokens before reuse.
• hooks/handoff-permission-request.py auto-approves safe PermissionRequest prompts for MCP calls and commands already covered by rules/default.rules, which helps current sessions continue when normal rules were not reloaded yet
• handoff service controls are approved for common app/process managers: pm2 start|stop|restart|reload, supervisorctl start|stop|restart, systemctl start|stop|restart|reload, service <name> restart, docker compose restart, docker compose up -d, direct docker|podman restart, and web-server reload commands for nginx/Angie/Apache/Caddy
• the safety hook still blocks destructive or mutating variants such as git reset --hard, git clean, force pushes, npm audit fix, go env -w, journalctl --vacuum-*, and mutating curl/wget requests
• persistent host/service mutations such as systemctl disable, systemctl mask, systemctl kill, Docker volume deletion, container prune, and host reboot/shutdown remain blocked or require explicit approval
• hook block messages include the read-only checks Codex should run next; selected Git cleanup/restore commands are allowed once after fresh git status plus git diff or git clean -nd review in the same working directory
• hooks/handoff-post-tool-use.py adds follow-up context after package installs, failed shell commands, git diff, and verification commands so Codex checks diffs/tests, maps results back to the task ledger, or fixes the concrete failure before repeating work
• MCP servers in this kit use default_tools_approval_mode = "approve" for handoff flow; agents must still keep database and local-machine MCP usage read-only unless the user explicitly asks for mutation
• templates/AGENTS.md includes handoff intake scoring and task-ledger rules. Non-trivial implementation work now assumes proactive subagent authorization by default, while questions, planning-only work, one-file fixes, and explicit inline-only requests stay inline. Approved Superpowers plans may use their documented worker split automatically, and implementation plans should include section review checkpoints plus a final code-review/verification pass.
• Claude Companion is wired into templates/AGENTS.md, agents_orchestrator, planning methodology, and code review skills as an advisory second-opinion reviewer. When claude and tmux are available, high-risk plans, completed Superpowers stages/worker groups, broad changes, data/security reviews, and release readiness checks should be invoked through $claude:* plugin commands such as $claude:code-review, $claude:security-review, or $claude:release-readiness-review with plan path, changed files, scope, verification, and concrete questions. The runner injects a compact senior-review methodology for each mode; Codex must still classify and verify every recommendation before applying it.

Optional prompt classification and LLM fallback:

mkdir -p ~/.codex/private
chmod 700 ~/.codex/private
printf 'OPENAI_API_KEY=...\nHANDOFF_CLASSIFIER_MODEL=gpt-5.4-nano\nHANDOFF_CLASSIFIER_LLM=auto\nHANDOFF_CLASSIFIER_TIMEOUT=6.0\n' > ~/.codex/private/handoff-classifier.env
chmod 600 ~/.codex/private/handoff-classifier.env

Use HANDOFF_CLASSIFIER_LLM=off for deterministic-only offline mode. auto and always both use the model whenever credentials are available; deterministic classification remains the fallback for API errors, timeouts, or unsupported structured output.

Dangerous mode:

./install.sh --force

Use it only when you intentionally want to replace managed files without backups.

Config, Plugins, MCP

The baseline config lives here:

templates/config.recommended.toml

The installer writes it here:

~/.codex/config.toml

The config enables:

• GitHub plugin entry: https://github.com/openai/plugins/tree/main/plugins/github
• bundled Superpowers fork plugin from https://github.com/VKirill/superpowers
• bundled Claude Companion plugin from plugins/claude-companion
• project docs discovery defaults
• agent concurrency defaults
• public remote docs MCP servers for Context7, Vue, Nuxt UI, and Nuxt

MCP Coverage

The starter kit separates portable MCP defaults from recommended local integrations. The local entries are part of the recommended full setup, but they stay commented in the public baseline because they need local daemons, paths, ports, bearer tokens, plugin state, or database credentials.

MCP or plugin	Status in this repo	Source	Why
github@openai-curated	enabled in templates/config.recommended.toml	https://github.com/openai/plugins/tree/main/plugins/github	GitHub repository, issues, pull request, and review workflow support
superpowers@codex-starter-kit	bundled local plugin, enabled by install.py through the local marketplace	https://github.com/VKirill/superpowers	forked planning, TDD, debugging, verification, and development workflow skills
context7	enabled in templates/config.recommended.toml	https://github.com/upstash/context7	public docs server, no local daemon required
vue-docs	enabled in templates/config.recommended.toml	https://github.com/joelbarmettlerUZH/vue-mcp	public Vue ecosystem docs
nuxt-ui-remote	enabled in templates/config.recommended.toml	https://github.com/nuxt/ui	public Nuxt UI docs
nuxt-remote	enabled in templates/config.recommended.toml	https://github.com/nuxt/nuxt	public Nuxt docs
serena	recommended local MCP; commented example in templates/config.recommended.toml; referenced by templates/AGENTS.md and agents	https://github.com/oraios/serena	semantic code navigation, references, and targeted edits
gitnexus	recommended local MCP; commented example in templates/config.recommended.toml; referenced by templates/AGENTS.md and many agents	https://github.com/abhigyanpatwari/GitNexus	code graph, impact analysis, route maps, execution flows, and repo context
postgres	recommended local MCP; commented example in templates/config.recommended.toml; referenced by templates/AGENTS.md and data/API agents	https://github.com/modelcontextprotocol/servers	local database inspection; tool calls are auto-approved for handoff, but agents should stay read-only unless explicitly asked to mutate data
open-design	recommended local MCP for design workspaces	https://github.com/nexu-io/open-design	local design artifacts, design-system context, and visual handoff
claude-mem	recommended local plugin/runtime for memory continuity	https://github.com/thedotmack/claude-mem	durable cross-session memory under ~/.claude-mem and mcp-search tools

templates/AGENTS.md intentionally tells Codex to use Serena, GitNexus, Context7, framework docs MCP, Open Design MCP, claude-mem, and database MCP when available. The baseline config enables only the portable public docs servers by default; recommended local integrations must be enabled after their daemons/plugins are installed.

For claude-mem, use its own installer/runtime flow, for example:

npx claude-mem@latest install

After setup, restart Codex and check what the runtime exposes:

codex mcp list

Manual runtime checks:

codex plugin marketplace upgrade
codex mcp list

Manual MCP add commands if you do not use the baseline config:

codex mcp add context7 --url https://mcp.context7.com/mcp
codex mcp add vue-docs --url https://mcp.vue-mcp.org/mcp
codex mcp add nuxt-ui-remote --url https://ui.nuxt.com/mcp
codex mcp add nuxt-remote --url https://nuxt.com/mcp

Custom Paths

Use a different Codex home:

./install.sh --codex-home /path/to/.codex

Use a different skills home:

./install.sh --skills-home /path/to/skills

Skip parts of the install:

./install.sh --skip-hooks
./install.sh --skip-rules
./install.sh --skip-skills
./install.sh --skip-agents
./install.sh --skip-config
./install.sh --skip-global-agents-md
./install.sh --skip-runtime-refresh

Repository Layout

codex-starter-kit/
├── agents/                     # Custom Codex subagents (*.toml)
├── skills/                     # Skills copied to ~/.agents/skills
├── hooks/                      # Shell safety hook and hook template
├── rules/                      # Codex command approval rules
├── templates/
│   ├── AGENTS.md               # Global project-agnostic Codex instructions
│   └── config.recommended.toml # Baseline ~/.codex/config.toml
├── prompts/
│   └── bootstrap-codex-starter-kit.md
├── scripts/
│   └── validate-pack.py
├── install.py
└── install.sh

Development

Validate the pack:

python3 scripts/validate-pack.py

Check installer dry run:

./install.sh --dry-run

Validate installed agents:

./install.sh --validate-only

What Not To Do

• Do not commit secrets, bearer tokens, or private MCP config.
• Do not put project-specific rules in the global AGENTS.md.
• Do not run --force if you need backups.
• Do not add local-only MCP ports to the public baseline config.

Keep project-specific instructions in each repository's local AGENTS.md.

GitHub Topics

openai-codex
codex-cli
codex-agents
codex-skills
agents-md
ai-coding-agents
mcp
subagents
developer-tools
coding-agent

Search keywords: OpenAI Codex agents, Codex CLI agents, Codex skills, Codex subagents, AGENTS.md template, Codex MCP setup, Codex hooks, AI coding agents, custom Codex agents and skills, Codex developer workflow, coding agent starter kit.

License

MIT