Awesome Codex Subagents: 136 Specialist AI Coding Agents Explained

By Beau Johnson··7 min read

Quick answer

VoltAgent's awesome-codex-subagents repo has 136 specialist Codex subagents in 10 categories. Each one is a TOML config with its own instructions, model routing, and sandbox permissions. You clone the repo, copy the TOML files into your Codex agents folder, and tell Codex which specialist to use for each task.

In this video I break down awesome-codex-subagents by VoltAgent. It's a GitHub repo with 136 specialized Codex subagents sorted into 10 categories, and it got more than 2,100 stars in its first week. It's for anyone using Codex, Claude Code, or OpenClaw who still runs one general-purpose agent for everything and wants to set up a team of specialists instead.

I recorded this in March 2026. These tools change fast, so check the repo and the current Codex docs before you set anything up.

What I cover in this video

  • 0:00: What the repo is, and why most people use Codex as a single generalist
  • 1:31: What a Codex subagent is and where its config file goes
  • 2:16: The core development team: backend, frontend, full stack, mobile, API design and more
  • 3:01: Language and framework specialists
  • 5:18: Model routing for lower costs, plus sandbox permissions
  • 6:50: Quality and security, DevOps, data and AI, docs, specialized domains, agent automation, and utilities
  • 10:39: How to install the agents and delegate work to them
  • 13:42: The companion repos and why specialist agents work better than one generalist

Why specialists beat one generalist

Most people open Codex or Claude Code, ask it to "build me a backend" or "review this code," and it gets the job done. But that's like asking one person to be a doctor, a lawyer, a plumber, and a chef at once. They can manage all of it, but a specialist will do each job better.

Codex lets you create custom agents that live in a specific directory on your machine. Each one is a small TOML config file in either your global agents folder or a project's agents folder. Codex can then call on that specialist. Every agent has its own instructions, its own model preference, and its own sandbox permissions. You're basically building a team: a frontend developer, a security auditor, an API designer, a database expert, and so on.

The 10 categories (the ones I walk through)

VoltAgent sorted the 136 agents into 10 categories. These are the ones I cover:

  • Core development: backend, frontend, full stack, and mobile developers, an API designer, a GraphQL architect, a microservices architect, a WebSocket engineer, and a UI-focused designer. The backend agent focuses on scalable APIs and server architecture. The frontend agent specializes in React and Angular. The mobile agent handles cross-platform work.
  • Language specialists: Python, JavaScript, TypeScript, Go, Rust, Java, C#, C++, Ruby, PHP, Swift, plus frameworks like Flutter, Django, Angular, and Next.js. These have deep framework knowledge. For example, the Django agent knows Django 4+ patterns, the Laravel agent knows Laravel 10+ conventions, and the Next.js agent knows the App Router. Generic advice leads to generic code. It's the same reason so many vibe-coded landing pages look alike.
  • Quality and security: a code reviewer, security auditor, performance profiler, test architect, accessibility expert, and tech debt analyst. Solo devs and small teams usually skip this work because they don't have time.
  • DevOps and infrastructure: a CI/CD engineer, Docker specialist, cloud architect, monitoring, and more.
  • Data and AI: a database expert, data pipeline builder (ETL and data quality checks), machine learning engineer (experiment tracking and model optimization), vector database specialist, and analytics dashboard builder.
  • Documentation and research: a technical writer, API docs generator, search specialist, and changelog manager. Documentation was always the thing I put off.
  • Specialized domains: fintech (financial compliance and PCI standards), healthcare compliance (HIPAA), e-commerce, real estate tech (MLS integrations), and game development.
  • Agent automation: an agent installer, an MCP developer, a workflow orchestrator, and an agent communication specialist. These are agents that help you build more agents.
  • Utilities: a Git specialist, regex expert, config manager, log analyzer, and migration specialist for database schema changes.

Model routing and sandbox permissions

This is the smartest part of the repo's design. Each subagent's config has a model field that sends it to the right model. Agents that need deep reasoning, like the security auditor and the architecture reviewer, run on GPT-5.4. Lighter scanning and synthesis jobs, like the search specialist and docs researcher, run on GPT-5.3 Codex Spark, which is faster and cheaper.

This is the same thing I keep saying about AI agents in general. Put your most expensive model on the coordinating agent. I run mine on Opus 4.6. Don't spend it on simple jobs like search or web scraping, or you'll hit your usage limits within an hour.

Each agent also has a sandbox mode. Reviewers and auditors are set to read-only, so they can look at your code but can't change it. Developers and engineers that need to write files get workspace write permissions. Not every agent can touch everything, and that's how you should think about agent permissions.

How to install and use them

  1. Clone the repo from the link above.
  2. Copy the TOML files where you want them. To make them available across all your projects, put them in your home directory under ~/.codex/agents. To use them in just one project, put them in that project's agents folder. If two agents have the same name, the project-level one wins.
  3. Restart your Codex session.
  4. Tell Codex which specialist to use. Codex doesn't spawn custom subagents on its own, so you have to name the agent in your prompt.

For example:

Use the security auditor to review this pull request.

Delegate this database migration to the database expert.

It works like project management: you assign each task to the right team member. If you use OpenClaw, you can give the repo to your OpenClaw agent and let it handle setup.

VoltAgent also made companion repos: Claude Code subagents, Claude skills, and agent skills that work across multiple platforms. So you can use the same specialist approach with other tools too.

Key takeaways

  • awesome-codex-subagents gives you 136 specialist Codex agents in 10 categories. Each one is a TOML file with its own instructions, model, and sandbox settings.
  • Model routing saves money. Put heavy reasoning on the strongest model and send lighter work to faster, cheaper ones.
  • Read-only reviewers and write-enabled builders keep agents from touching things they shouldn't.
  • Codex won't pick a custom subagent for you. Name the specialist in your prompt.
  • The pattern of specific instructions, model routing, and sandbox permissions works in any tool, whether that's Codex, Claude Code, OpenClaw, or Cursor. I learned the hard way that one agent doing everything keeps failing.

FAQ

What is awesome-codex-subagents?

It's a GitHub repo from VoltAgent with 136 specialized Codex subagents sorted into 10 categories. Each agent is a TOML config file with its own instructions, model preference, and sandbox permissions.

Where do I put Codex subagent TOML files?

To use them across all your projects, put them in your home directory under ~/.codex/agents. To use them in one project, put them in that project's agents folder. Project agents win over global ones if two have the same name.

Does Codex automatically use custom subagents?

No. Codex doesn't spawn custom subagents on its own, so you name the one you want in your prompt, like "use the security auditor to review this pull request."

Which models do the Codex subagents use?

Each agent has a model field in its config. Deep-reasoning agents like the security auditor run on GPT-5.4, and lighter agents like the search specialist run on the faster, cheaper GPT-5.3 Codex Spark.

Chapters

  1. 0:00What is awesome-codex-subagents?
  2. 1:30Why specialists beat generalists
  3. 3:00Core development agents
  4. 4:30Language specialist agents
  5. 6:00Smart model routing (cost optimization)
  6. 7:00Security and quality agents
  7. 8:30DevOps, data, and domain specialists
  8. 10:00How to install and use them
  9. 11:30Why this pattern matters for everyone

Transcript

Read the full transcript

0:00There's this GitHub repo we need to talk about with over 2100 stars in just one week. This repo replaces your entire dev team in just seconds. We're going to be talking about it and how we can use it and implement it in our open claw setups to help buy back some of your time and hit the ground running if you're just getting going with AI agents. So, like I said, this thing just hit 2100 stars in a week and there's a good reason for it. You know, if you've been using Codex at all or Claude code and or open claw, you know, this is very powerful. You know, you open it up, you give it a task, it writes code, it does it in a sandbox environment, and then that's great. But, the thing that most people don't realize and most people are using Codex basically like a single generalist, same thing with Claude code. You can say

0:46like, "Hey, you know, build me a back end or you know, review this code, fix my CES, things like that, right?" And it does it, gets the job done. But, you're basically asking one person to be a doctor and a lawyer and a plumber and a chef all at the same time. They can do all of those things, but wouldn't you rather have a specialist? That's where these AI agents are pretty much drag and drop. 136 of them to be exact. I have the repo linked down below. You got to check it out. So, this is exactly what the repo gives you. Basically, 136 specialists. I know that sounds overwhelming, but I want to break down the key ones in this repo. So, you got sub agents. Basically, each one is designed to do one specific type of development task very, very well. What exactly is a sub agent because there's a lot of confusion around this whole AI

1:31agent space. If you're not familiar, Codex has this feature where you can create a custom agent that live in a specific directory on your machine. You put in a little tunnel file config and basically your global agents folder, your project agents folder, and now Codex can call on that specialist whenever it needs to. So, you can think of it like building a team, basically. Your front end developer, your security auditor, your API designer, basically a database expert, things like that, right? Each one has their own instructions, you know, their own model preferences, their own sandbox permissions. But, this is what Vault agent built. 136 categorized sub agents into 10 categories. I'm going to walk you through some of them because this is where it gets more interesting is

2:16category one is the core development team. This is where your bread and butter is like you have a back end developer agent, a front end developer agent, a full stack developer, a mobile developer, an API designer, and a GraphQL architect, a microservices architect, a web socket engineer, basically a engineer and designer that focuses on UI as well. So, right there is like 10 specialists just for building things. The back end developer knows about scalable APIs, the server architecture, the front end developer specializes in React, Angular. The mobile developer does cross-platform work. Each one has deep specific instructions about how to approach that type of work. The second category is

3:01more language specialists and this one's a huge one because if you're working in a specific language, you want an agent that actually knows that ecosystem deeply. So, they have agents specific for Python, JavaScript, TypeScript, Go, Rust, Java, C#, C++, Ruby, PHP, Swift, all of them, guys. Flutter, Django, Angular, Next.js. It They the It's a little bit overwhelming this what this repo offers. That's why we're making a video about it. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full open claw course, and four live bootcamps every single week

3:46where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you can book a call with me, we can share screens, and I can help you get Claude code or open claw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Each one is not just a generic coder that happens to know the language. Each one has deep framework knowledge

4:31specific to that stack. So, let's say your Django developer knows Django 4+ patterns. Your Laravel specialist knows Laravel 10+ conventions. Your Next.js expert knows app router. That specificity matters because generic advice leads to generic code. We've all been there before when you vibe code a landing page, it looks like every other vibe coded landing page in existence. That's why it's important to have specialized agents with specialized skills and knowledge to be able to stand out and build a unique product. Here is something that's really smart about what they did with model routing. Each sub agent has model field in its config file that routes it to the right AI model automatically. So, your security auditor

5:18and your architecture reviewer, you know, the ones that need deep reasoning, those run on GPT-54. And that's the heavy hitter. But, your search specialist and your docs researcher, and the ones doing more of the lighter work like scanning and synthesis, those run on GPT-53, Codex Spark, which is much faster and cheaper. So, you're not burning most of the expensive model on simple task. You're running routing intelligent. That is cost optimization built right into the agent design. And if you've been watching my videos, you know I talk about this all the time. The number one people mistake they make when getting AI agents up and running, you know, they get the most expensive model, which you should for the coordination agent. You know, I I run it on Opus 46. But, you

6:05don't need Opus 46 for simple things, you know, like searching and web scraping. You're going to blow through your usage limits in an hour. You know, and the same principle applies here is like figure out what models can be delegated to handle the simpler tasks that a you know, more more token efficient LLM can do. You know, they also thought about this when they built the repo. You know, they thought about security, which I always appreciate. Each sub agent has basically a sandbox mode setting. Reviewers and auditors are set to read only. They can only look at your code, but they can't modify anything. Developers and engineers that actually need to write files, those get workspace right permissions. So, you have got this nice separation of concerns where not every agent can touch everything. That is exactly how you

6:50should be thinking about agent permissions. So, I'm going to talk about some of the categories that really stood out in this huge repo. Category four is quality and security. You have a code reviewer, a security auditor, a performance profiler, a test architect, an accessibility expert, and a tech debt analyst. So, if you think about that for a second, you could have a dedicated agent whose entire job is to look at your code base and find tech debt. Or one that specifically audits security vulnerabilities. Or one that checks your app for accessibility compliance. These are the things that most solo devs and even small teams just skip because they don't have the time. Now you have a specialist for each one sitting in a tunnel file ready to go. Category five

7:35is the DevOps team and infrastructure. CI/CD engineer, you know, a Docker specialist, a cloud architect, monitor, and all of that. You know, if you're deploying apps and managing infrastructure, you basically now have a team of agents that know exactly how to handle all of that. And the fact that they separated this into its own category tells you how important infrastructure automation is becoming. The sixth category is data and AI. So, the database expert, data pipeline builder, the machine language the machine learning engineer, vector database specialist, analytics dashboard builder, you know, just the fact that they have a vector database specialist tells you where the industry is headed. Everyone is building with embeddings and semantic search now, and having an agent that specializes in that is generally

8:21useful. The machine learning engineer knows about experiment tracking and model optimization. The data pipeline builder understands ETL workflows and data quality checks. Category seven is documentation and research. So, you have a technical writer, an API docs generator, a search specialist, a change log manager. These are all tasks that nobody wants to do, but everyone needs it to be done. So, having an agent that specifically writes clean API documentation or manages your change log, that saves hours every single week. I know from building out our mission control here that documentation is the thing that I was always putting off. But, having an agent dedicated to this will help greatly. Category eight is specialized domains. So, this is where

9:06it gets more interesting for people building actual real products. Fintech engineer, healthcare compliance specialist, e-commerce developer, real estate tech developer, game developer. These agents know domain specific rules and regulations. So, like your Fintech agent understands financial compliance and PCI standards. Let's say your healthcare agent knows about HIPAA requirements. Your real estate tech developer understands MLS integrations. These are not generic coders now. We're past that. These are agents with real domain expertise baked into their instructions. So, if you're going to be building one of in one of these industries, this alone is worth cloning this repo. Category nine is agent automation. So, which is where the real meta is at. You know, you have got an agent installer, an MCP developer for

9:53building model context protocols, a workflow orchestrator, and an agent communication specialist. So, you have agents that help you build more agents. That is the pattern we are seeing everywhere now. The tools are building tools. The category 10 is utilities, so you have a Git specialist, a regex expert, config manager, log analyzer, migration specialist. These are all little tasks that eat up your entire day if you let them, and we're all trying to optimize to get our time back. I am at least. So, having a dedicated regex expert that you can delegate to and instead of spending 30 minutes on Stack Overflow trying to figure out the right pattern, having a migration specialist that handles database schema changes cleanly, that's time back in your pocket. So, how do you actually use all

10:39this? Like, if you're a beginner watching this, this probably is overwhelming. I'm only making this video because AI agents is the hot topic now, and a lot of people get Open Claw going or Codex going or Claude Code going, and they're just working at it out of the box, default level. Your agents need domain-specific knowledge. They need skills to be able to execute tasks and get good quality output. You know, it's really dead simple. You go to the link in the description, you clone the you clone the repo, you copy the TOML files, and you put them in the Claude Code, you put them in the Codex, give them in the Open Claw. If you're struggling to keep up with content, what I'm about to save you about 40 days worth of work. [music] I built something called Content Machine. It's 10 AI agents that run on the Open Claw orchestration, and they handle everything. Scripts, thumbnails, X

11:25posts, [music] blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up and [music] the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. It works for any niche, fitness, [music] finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns

12:11how you talk and it writes [music] so it doesn't sound like AI slop. $97 one time, it's not a subscription. I'll put the link down below and you'll thank me later. You know, if you want them available globally across all your projects, you put them in your home directory under .codex/agents. Likewise, just give the repo to your Open Claw agent and let them do the rest. So, if you want them in specific to one project, you put them in that project's So, project agents take priority over global ones if there's a naming conflict, you know, which is pretty smart design. Then you just restart your session and you can start delegating to them explicitly, you know, in your prompts. One thing to note is that Codex does not auto-spawn custom sub-agents,

12:57so you have to tell it which specialist to use. So, you would say something like, use the security auditor to review or pull or use the do this pull request or delegate this database migration to the database expert. It's like being a project manager and assigning tasks right to the team member. Here's what makes this more powerful and why it connects to what I've been doing. You know, I run the entire content operation in my business of specialized AI agents on the Mac Mini over here, and I've got the Rizza handling video outlines, Inspector Deck handling everything on X, Ghostface doing intel research, Method Man doing clips, Gizza running the security. You know, each one has specific jobs and specific personality. That is the pattern that works. Specialization beats generalization every time. I learned that the hard way by trying to make one agent to do

13:42everything and it would just fail over and over again. This repo proves that the developer community is figuring this out at scale. You know, 2,100 stars in a week doesn't happen by accident. People are realizing the way to get the most out of AI coding tools is not to use one agent for everything. It is to build a team of specialists and orchestrate them intelligently, but not one doing everything. And that's what Vault Agent also did, and I think that's really smart. They built a companion of repos, and they have awesome Claude Code sub-agents for Claude Code users, awesome Claude skills for Open Claw users, and awesome agent skills that cover skills across multiple platforms. So, regardless of what tool you're using, they're trying to give you that specialist approach that tells you and tells me that the team understands, you

14:28know, the future is not about one platform winning. It's a pattern of specialized agents working for you everywhere 24/7. The thing I want to you to take away from all this is that the era of generic AI assistant is completely gone. We are moving into the era of AI teams. Whether you're using Codex, Claude Code, Open Claw, Cursor, whatever you're using, the builders who figure out how to create specialized agents and orchestrate them are going to absolutely crush the market because you are not limited by your own time anymore. You are limited by how well you set up your team. And if you want to learn how to actually build out these agent teams for your business, if you want to go from using AI as a tool to using it as a real competitive advantage, that's exactly what we teach inside Shipping School. We are on day 13 of launching the community, and we have

15:14134 members. We've got builders in there right now running agent teams for their businesses, real people getting real results. We have people that are non-technical getting Open Claw installed and launching their landing page within a day. Bunch of people from all over the world, and we're really focused on community first, which is why we do six boot camp calls every single week, and everything's recorded. We do have courses in there as well, but we're focused on human-to-human interaction. So, like I said, I'll put that repo and the link to the community down below. Check it out. You know, we got awesome Codex sub-agents on GitHub and many other different types of repos that you could plug and play to get your AI agent orchestration going. You know, even if you don't use Codex, for example, the patterns they use and the way they structure agents with specific

15:59instructions and model routing and sandbox permissions, you can apply those same patterns to whatever tool you're using. The framework doesn't matter here. The pattern does. I just hope you're taking advantage of all this AI agent news and applying it to your setup because that's what makes the difference. If you haven't subscribed already and you like AI news just like this, make sure you subscribe, hit the notification bell. We literally release four to six videos every single day. Have a blessed one. We'll see you in about an hour.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading