OpenAI AgentKit vs AgentKit (agentkit.best): Key Differences (2026)
"OpenAI AgentKit" (launched at DevDay, October 2025) and "AgentKit" (agentkit.best) share a name but are two completely different products. The first is OpenAI's platform for building agent apps from scratch - Agent Builder, ChatKit, Connector Registry, and Evals - and two of those four pieces, Agent Builder and Evals, are on a wind-down path (shutting down 2026-11-30). The second is a paid kit that installs prebuilt skills/subagents directly into the Claude Code or Codex you already use. This page puts both side by side, plus a router to pick the right one in 30 seconds.
- Launch date, product components, and the wind-down timeline were cross-checked against OpenAI's official documentation at the time of writing; verify the live docs before you depend on them, since the roadmap moves fast.
What is OpenAI's AgentKit?
OpenAI AgentKit launched at DevDay on 2025-10-06, announced by Sam Altman. It's a four-part toolkit for developers and enterprises to build agent apps from scratch on OpenAI's own platform: Agent Builder (a visual drag-and-drop canvas for agent workflows, in beta), ChatKit (an embeddable chat interface, GA), Connector Registry (secure connections to internal and third-party systems, beta), and Evals (performance measurement, trace grading, automated prompt optimization).
In short: it's for taking an agent app from prototype to production on OpenAI's infrastructure - not something you install into Claude Code or Codex. Per TechCrunch and OpenAI's own announcement page, several launch partners had already scaled real agents on the platform before it went public.
What is AgentKit (agentkit.best)?
By contrast, AgentKit (agentkit.best) has nothing to do with OpenAI. It's a paid CLI called ak that installs prebuilt skills/subagents/workflows directly into Claude Code and Codex, which you already have installed - not a platform for building an agent from zero, but an extension layer for the coding agent you already use daily (for example ak kit init engineer --target codex).
It fits developers already comfortable with Claude Code/Codex who want a prebuilt kit instead of assembling a review-ship workflow by hand. Install is a single script (curl for macOS/Linux, irm for Windows) with no Node or Bun dependency, then ak login to activate before running a kit. To be direct: this is a paid add-on, while Claude Code and Codex themselves remain free or bundled into a ChatGPT plan. AgentKit also has an experimental Cursor adapter, but kit commands aren't fully functional there yet - no status line, no plan tool, so don't assume full parity. The product is built primarily for Claude Code and Codex. For the full picture, read the complete AgentKit review.
Side-by-side comparison
Line up both products by criterion, so you don't have to read two separate pages to get the picture:
| Criterion | OpenAI AgentKit | AgentKit (agentkit.best) |
|---|---|---|
| Maker | OpenAI | agentkit.best (third party, unaffiliated with OpenAI) |
| Launched | 2025-10-06 (DevDay) | Active (2026) |
| What it is | Agent-app-building platform: Agent Builder + ChatKit + Connector Registry + Evals | ak CLI that installs prebuilt skills/subagents/workflows |
| Runs where | OpenAI's own cloud platform | Inside the Claude Code or Codex you already have installed |
| Who it's for | Devs/enterprises building agent apps from scratch | Devs already using Claude Code/Codex who want a prebuilt kit |
| Cost | OpenAI API/usage pricing | One-time paid purchase (Engineer Kit at the listed price at time of writing - check the live pricing page) |
| Current status | Agent Builder + Evals shutting down 2026-11-30; ChatKit + Connector Registry remain | Actively maintained |
The confusion is easy to explain: every article currently ranking for "OpenAI AgentKit" assumes it's the only AgentKit in existence, and none of them are dated after October 2025 despite the June 2026 wind-down news existing. If you searched hoping to confirm the kit you already installed in Claude Code or Codex, you landed on pages describing a completely different OpenAI product instead.
A third name - AGENTS.md is not a product
AGENTS.md is a configuration file, not a product or a company. Codex (and many other coding agents) read this file from disk to learn your project's conventions - a completely different concept from either "AgentKit" above. Because the names sound alike, some people mistype "agents.md" or "agent kit" and land on one of the two pages above by accident. To see exactly how AGENTS.md differs from CLAUDE.md, see AGENTS.md vs CLAUDE.md.
Which one do you actually want?
I've (Jasmine) seen this exact mix-up show up repeatedly in reader questions: "Is the AgentKit I installed in Claude Code the same one I just read about in the news?" No - and the answer depends on what you're actually trying to do.
- Building a custom agent app from scratch on OpenAI's own platform (drag-and-drop canvas, ChatKit embedded in your product, Connector Registry, Evals) - something your company will host and ship as its own product? You want OpenAI AgentKit - see OpenAI's official docs at openai.com. Nothing here installs into a coding assistant.
- Already using Claude Code or Codex day to day and want a prebuilt kit (skills/subagents/workflows) instead of assembling a review-ship process yourself? You want AgentKit (agentkit.best) - see using AgentKit inside Codex or whether AgentKit's pricing is worth it. This is the one people mean when they ask "how do I get AgentKit working in my terminal."
Already on Claude Code or Codex and just want the prebuilt kit instead of assembling your own? AgentKit (agentkit.best) installs skills/subagents/workflows directly into the coding agent you already have - it is not a build-your-own-agent platform like OpenAI AgentKit. Priced per the live pricing page at time of purchase, often with a first-purchase discount.
Is OpenAI's AgentKit going away?
Partly, yes. On 2026-06-03, OpenAI announced it's winding down Agent Builder and Evals: Evals becomes read-only starting 2026-10-31, then both (Agent Builder and Evals dashboard + API) shut down completely on 2026-11-30. ChatKit and Connector Registry are unaffected and continue operating normally.
To be fair about it: this isn't OpenAI killing "AgentKit" wholesale - only two of the four components are shutting down, and ChatKit remains a real GA product. And none of this touches AgentKit (agentkit.best) at all - the two products are entirely independent, so one winding down has zero effect on the other. Source: OpenAI's official deprecations page (re-check it, since dates can shift).
Frequently asked questions (FAQ)
Is OpenAI's AgentKit the same as AgentKit (agentkit.best)?
No. Different makers, different purpose. OpenAI AgentKit is a platform for building agent apps (Agent Builder, ChatKit, Connector Registry, Evals), made by OpenAI. AgentKit (agentkit.best) is a paid, third-party kit that installs prebuilt skills/subagents into Claude Code and Codex - two unrelated products that just happen to share a name.
What happened to OpenAI's Agent Builder?
OpenAI announced on 2026-06-03 that it's winding down Agent Builder and Evals. Evals goes read-only starting 2026-10-31, and both shut down completely on 2026-11-30. ChatKit and Connector Registry are unaffected.
Does AgentKit (agentkit.best) work in Cursor?
Not fully. It's built primarily for Claude Code and Codex; there's an experimental Cursor adapter, but kit commands aren't fully functional there yet - don't assume full parity.
Is AgentKit (agentkit.best) free?
No - it's a one-time paid add-on. Claude Code and Codex themselves remain free or bundled into a ChatGPT plan; AgentKit just adds a prebuilt kit on top.
What is AGENTS.md?
A configuration file that Codex and many other coding agents read to learn your project's conventions - not a product, and unrelated to either "AgentKit" above.
I use Claude Code or Codex - which AgentKit should I use?
AgentKit (agentkit.best) - it's the one that installs directly into the Claude Code or Codex you already use. OpenAI AgentKit is a different platform for building your own agent app on OpenAI's infrastructure, not something you install into your coding agent.
Conclusion
They only share a name - nothing else. If you're building your own agent app on OpenAI's platform, read OpenAI AgentKit's official docs. If you're already using Claude Code or Codex and just want a prebuilt kit instead of assembling your own process, that's AgentKit (agentkit.best). And if you spot "AGENTS.md" somewhere - that's a configuration file, not either product above.