How to Use AgentKit With OpenAI Codex (2026 Step-by-Step)

- Commands and prices below were cross-checked against the official docs (docs.agentkit.best, beta channel) at the time of writing, 2026-08; verify the live docs before depending on them, since the beta channel moves faster than a GA release.

Three things named "AgentKit"/"Codex" - read this first

Before you type a single command, separate three names that keep getting conflated. Mixing these up means buying the wrong product at the wrong price, not just a wording mistake:

This is why searching "use AgentKit with Codex" or "AgentKit pricing" surfaces OpenAI's own Agent Builder results in both English and Vietnamese - wrong product, wrong price, wrong install path. This article is only about row two above: agentkit.best. For the full breakdown so you don't buy the wrong thing, see OpenAI AgentKit vs AgentKit (agentkit.best) - the full comparison.

What you need before you start

A five-minute checklist - read the whole thing before you type your first command:

• Codex CLI, or a ChatGPT plan that includes Codex, already installed and signed in with your OpenAI account.
• An AgentKit license you've already purchased - Engineer $99 or Bundle $149, not a free tier.
• A terminal open in the exact project directory you want the kit installed into.
• About 5 uninterrupted minutes.

The detail people skip: logging into AgentKit does NOT sign you into Codex, and vice versa - these are two fully separate auth systems, so you'll log in twice for two different things, and you'll need to re-authenticate each one separately if you switch machines. Don't get to ak kit init and discover Codex was never signed in, or the other way around - checking both first saves a pointless debugging loop. Haven't installed Codex yet? Read how to install the Codex CLI first. Not sure what Codex even is, or how it differs from Claude Code? See what OpenAI Codex is.

Step 1 - Install the ak CLI

AgentKit runs through a single native CLI binary called ak - no Node, no Bun, no extra runtime to install first. Pick your OS:

# macOS / Linux
curl -fsSL https://agentkit.best/install.sh | sh

# Windows (PowerShell)
irm https://agentkit.best/install.ps1 | iex

The installer detects your OS and architecture automatically. Once it finishes, verify with two commands:

ak --version
ak doctor --json

ak doctor --json prints your install health as JSON - run it any time something feels broken, including in later steps. Because it's a native binary, this install path also works cleanly in sandboxed/CI environments with no Node available - a real difference versus kits shipped as npm packages.

Step 2 - Sign in and confirm your license

Sign in with email or an API key, whichever's more convenient:

ak login --email [email protected]
# or
ak login --api-key ak_live_...

Then confirm you're on the right account with the right license before installing anything - this is the step most people skip, then discover the license never actually activated, or they're logged into a different email than the one used at checkout:

ak whoami
ak licenses

ak licenses lists exactly what you own (Engineer, Marketing, or Bundle). If the list comes back empty even though you're sure you paid, go check your order confirmation email or your account page on agentkit.best before continuing - don't install blind and then wonder why it errors on a missing entitlement.

Step 3 - Install the kit into Codex

This is the central command of the whole article - the one no page anywhere, EN or VI, currently writes out in full:

ak kit init engineer --target codex

By default this installs at project scope - the kit only applies inside the current project directory, not everywhere. If you want it available in EVERY project without reinstalling each time, add the global flag:

ak kit init engineer --target codex --global

Before it writes anything to disk, ak shows a confirm/preview screen: the kit name, the output location, the install mode (project or global), the scope, and the exact list of files it's about to write. Read this screen before you type "y" - it's the moment you know precisely what the kit will touch, instead of guessing after the fact and digging through what got written.

Hands-on note: running this the first time, the confirm screen laid it all out plainly - Engineer kit, writing to .agents/skills, project scope, and a concrete list of files about to be created. Nothing hidden behind it; hit confirm and you know exactly what got touched.

Don't have a kit yet? You need an Engineer or Bundle license before the command above will actually run. Check current pricing and any active deal, then come back to this step.

Get the AgentKit Engineer Kit →

Step 4 - Run your first skill with $ak:cook

Codex only recognizes the skill after you open a fresh Codex session in the same project you just installed into - it needs a reload to re-scan for skill discovery. Don't keep typing in the old session and wonder why nothing happens.

In the new session, invoke the skill with Codex's $ak: syntax:

$ak:cook Add a health-check endpoint

Swap the sample task for your own - that's just an illustrative example of the calling syntax, not a fixed formula. The detail worth remembering is the syntax split between the two runtimes: Codex uses $ak:, while Claude Code uses /ak: - same kit, same skills, just a different invocation character depending on where it's running. Don't type /ak:cook inside Codex and assume the kit is broken; that's almost always a syntax mismatch, not an install failure.

Hands-on note: calling $ak:cook in a fresh Codex session, the skill was picked up immediately with no extra configuration - matching the docs' description that discovery needs a new session after install.

Where AgentKit behaves differently inside Codex vs Claude Code

This is the genuinely new information here - no other page currently writes this out. AgentKit doesn't run identically on every runtime, and Codex has specific limits versus Claude Code, spelled out in the runtime-adapter docs (docs.agentkit.best/en/beta/concepts/runtime-adapters, accessed 2026-08):

Short version: skills, rules, and agent dispatch run fully in Codex - that's the whole mechanism behind calling $ak:cook. Hooks are partial, and kit commands (the kit's own slash-style commands, distinct from skills) plus the status line simply don't exist on Codex. If a workflow you rely on leans heavily on kit commands or a hook outside the supported group, expect that piece to silently not run on Codex even though it works fine on Claude Code - that's not a bug, it's the adapter declaring its boundaries up front, not something you discover the hard way after depending on it.

If you're used to Claude Code and Codex is new, don't be surprised when a few things "disappear" - they were never supported on Codex to begin with, not broken by your install.

A concrete example: if a workflow you're used to on Claude Code includes a step that fires through a kit command, or depends on a hook outside the supported group, expect that specific step to quietly not run on Codex while the core skill layer ($ak:cook and the rest) keeps working fine. The safe move is to test right after install rather than assume identical behavior just because it's the same kit.

Free Codex vs paid AgentKit - the honest boundary

Said plainly up front: you are not paying to "use Codex." Codex CLI, Codex cloud, and even goal mode /goal are free or already covered by the ChatGPT plan you pay for (Free/Go/Plus/Pro/Business). AgentKit is a layer of prebuilt skills, agents, and workflows you install on top of that same free Codex engine - you're paying to skip writing the skills yourself, not to unlock anything inside Codex.

Prices above reflect the pricing page at the time of writing (2026-08); the site frequently runs a -20% deal through the ref link (Engineer at roughly $79.20) - check the current banner before you pay, don't treat any single number as fixed since the deal can change. There's a money-back guarantee, but the site doesn't publish a specific window - confirm the exact terms at checkout instead of assuming "14 days" or "30 days." For the full product picture and an hourly-ROI breakdown before buying, read the full AgentKit review or the detailed pricing at AgentKit pricing, is it worth it - those two go deep on valuation; this article only covers the free-vs-paid line clearly enough to decide before you install.

Troubleshooting

• Suspect a broken install: run ak doctor --json and read the error section of the JSON before reinstalling from scratch.
• Want to check kit/skill versions: ak versions compares against the latest release, useful if you suspect the kit is out of date.
• $ak:cook isn't recognized: almost always means you haven't opened a fresh Codex session in the exact project directory the kit was installed into - close the old session, reopen in the right directory, try again.
• Installed at project scope but need it elsewhere: that's expected behavior, not a bug - project scope only applies where you installed it; rerun the command with --global if you want it everywhere.
• Not sure the license is still valid: ak whoami plus ak licenses confirms the account and which kits it actually owns.
• Confirm screen shows something you didn't expect (wrong scope, wrong output path): cancel out of it, double-check whether you meant to add --global, and rerun - it's cheaper to cancel a preview than to clean up files written to the wrong place.

Frequently asked questions (FAQ)

Is AgentKit free when used with Codex?

No. Codex (CLI/cloud/ChatGPT) is free or already covered by the plan you pay for. AgentKit is a paid kit installed on top - Engineer $99, Bundle $149, the same price for any target including Codex.

What's the difference between /ak: and $ak:?

Same kit, same skills, just a different invocation syntax per runtime. Claude Code uses /ak:cook; Codex uses $ak:cook. Typing the wrong character is the single most common reason a skill "doesn't run."

Do I need Claude Code too, or can I install AgentKit for Codex only?

Codex alone is enough. ak kit init engineer --target codex installs independently for Codex - Claude Code isn't required alongside it.

Is this the same as OpenAI's own AgentKit?

No. The AgentKit in this article is agentkit.best, a third-party product driven by the ak CLI. OpenAI AgentKit is OpenAI's own Agent Builder/ChatKit - a completely different product that just shares the name.

Does AgentKit cost more for Codex than for Claude Code?

No difference. The same license, at the same price ($99 per kit or $149 for the Bundle), applies to whichever --target you install - Codex isn't priced separately or higher.

Can I use one license across Claude Code and Codex?

There's no public confirmation that a single license allows unlimited simultaneous targets - licenses are managed by device activation, and --target is just an install-time parameter for where the kit lands. If you plan to run both runtimes on the same license, confirm the activation limits at checkout rather than assuming.

I haven't installed Codex yet - can I install AgentKit first?

Not really worth it. ak kit init … --target codex only writes skills into a directory Codex reads (.agents/skills); with no Codex installed, there's nothing there yet to run it. Install Codex first via the Codex CLI install guide, sign in, then come back to the steps above.

Does the confirm screen actually let me cancel before anything is written?

Yes - that's the point of it. ak kit init shows the kit name, output location, install mode, scope, and file list before writing anything; you can back out at that screen if the scope or destination looks wrong, instead of finding out after the fact.

Conclusion

The whole real path fits in four steps: install ak, ak login, ak kit init engineer --target codex, open a fresh Codex session, then call $ak:cook. Codex stays free or already covered by the ChatGPT plan you pay for; AgentKit is the extra cost for a prebuilt skill/agent layer, not a fee to "unlock" Codex. Worth buying if you already use Codex daily and don't want to write skills from scratch; not urgent if you're still getting used to plain Codex.