Product details and OpenAI transition dates checked November 3, 2026. Originally published June 22, 2026.
Agentkit at agentkit.ai and OpenAI's AgentKit are separate products. Use Agentkit if your job is to put a chatbot that answers from your website or documents in front of visitors. Use OpenAI's developer tools if you need to build an agent into your own application and control its code, server, and tools. This comparison is published by Agentkit; the similar names do not imply an affiliation.
The distinction matters more in 2026 because OpenAI is retiring its visual Agent Builder while keeping ChatKit available. If you are choosing a new development path, check that transition before starting a build.
Choose by the job you need done
| Your job | Where to start | What you need to do |
|---|---|---|
| Answer questions from your public website or documents on an existing site | Agentkit | Add sources, review answers, enable the website channel, and follow the embed guide. |
| Build a custom agent inside an application you operate | OpenAI Agents SDK, with ChatKit if you need its chat interface | Write and run the agent in your application, connect its tools, and host the server-side integration. |
| Replace an existing OpenAI Agent Builder workflow before its scheduled November 30, 2026 shutdown | OpenAI's migration guide | Export its Agents SDK code, then choose an application using the SDK or a ChatGPT Workspace Agent. Check what must be recreated and test the replacement. |
What you get with Agentkit
Agentkit is the hosted chatbot service on this site. You create an agent, add website pages or documents, test its replies, and publish it to a site you control. The dashboard and website widget are part of that service. If you need a floating chat bubble, use Channels → Website embed → Chat widget. For an inline chat panel, enable the separate Iframe option and follow the iframe guide.
For a new installation, copy the complete generated code from Channels and use the website embed guide to test the public page.
Check the current plan limits before treating a pilot as a production capacity estimate. You still own source review, answer testing, website access, and any privacy review your use case requires. Agentkit offers configured actions through supported integrations, but it does not give you the agent's source code or a runtime to host yourself. For custom workflow code and direct control over tools, use a developer stack. If a hosted website chatbot fits, create an agent and test it with a small source set you have reviewed.
What OpenAI's AgentKit means now
OpenAI's original AgentKit walkthrough grouped Agent Builder, ChatKit, and Evals as developer building blocks. The walkthrough is archived and warns that parts may be outdated, so use the current guides below for a new build.
OpenAI's deprecations page says:
- Agent Builder was deprecated on June 3, 2026 and is scheduled to shut down on November 30, 2026.
- Existing Evals became read-only on October 31, 2026; the Evals dashboard and API are scheduled to shut down on November 30, 2026.
- ChatKit remains available as of this November 3 documentation check.
For a new coded agent, OpenAI's Agents SDK guide describes a runtime in your own application. Your server owns deployment, storage, tools, and approvals. OpenAI's ChatKit guidance directs new chat implementations to its advanced integration with your own server-side agent. The older path that connects ChatKit to an Agent Builder-hosted workflow is for existing workflows during the transition.
If you already use Agent Builder, follow OpenAI's migration guide. The export gives you TypeScript or Python Agents SDK code; it does not convert the workflow graph or guarantee the same behavior. The guide also offers a ChatGPT Workspace Agent path for Business, Enterprise, or Edu workspaces with access and creation permission. Test control flow, triggers, tools, permissions, and representative inputs before replacing the old workflow.
Agentkit can replace a Builder workflow whose actual job is answering from approved website or document content on a site you control. A workflow that depends on custom code, multi-step tool control or your own runtime needs a developer path; Agentkit's supported actions are not a drop-in conversion for it.
For one existing workflow, make a replacement record before you switch traffic:
| Check | Record before export | Verify in the replacement |
|---|---|---|
| Entry and result | A representative input, expected answer or action, and the user who can start it | The same user can start it and gets the intended result |
| Tools and permissions | Each connected tool, account, scope, and approval step | The connection works with the intended account; an unauthorized user cannot run the action |
| Branches and failures | A case that takes each important branch, plus a tool failure | The replacement follows the intended branch and reports the failure truthfully |
| Conversation state | What the workflow needs from an earlier turn | A resumed conversation retains only the state it needs |
Compare the replacement with those saved cases before retiring the original. A code export is a starting point, not a verified migration. OpenAI says strongly deterministic workflows may not migrate faithfully to a Workspace Agent.
Compare the work and the bill
Agentkit's hosted plans and allowances cover the chatbot service. OpenAI's API pricing page lists model token charges and separate charges for some tools; your team also has to build and operate the application. Compare the allowances and likely usage, then add your own hosting and engineering costs. Neither page alone gives a total cost for your use case.
