Quick comparison (decision matrix)
| Dimension | Apps SDK | AgentKit | Responses API |
|---|---|---|---|
| Primary outcome | ChatGPT-native app with UI inside the chat | Agentic workflow you host & embed | API for model calls & built-in tools |
| Where it runs | Inside ChatGPT (web & mobile) | Your product (web/mobile/desktop) | Your servers & client apps |
| Distribution | In-chat suggestions; upcoming app submissions & directory | Your channels (website/app), no ChatGPT directory | Your channels; no ChatGPT surface |
| UI layer | Native components rendered in chat | ChatKit embeddable UI (you host) | You build/own the UI |
| Logic & tools | Via MCP tools exposed by your server | Agent graphs via Agent Builder; tools/connectors | Tool use via Responses tools (web/file/computer) |
| Testing & quality | Developer Mode in ChatGPT | Evals + Builder previews | Your tests + Evals (optional) |
| Best for | Reaching ChatGPT users; guided flows, lead gen, support | Complex, multi-step agents in your app | Programmatic use, automations, APIs |
When to choose each (clear, verifiable criteria)
Choose Apps SDK if…
- You want a ChatGPT-native experience with embedded UI (lists, forms, maps, previews) and MCP access to your data.
- You value in-ChatGPT distribution (suggestions, future directory) and can operate within the current preview + Developer Mode constraints until submissions open later this year.
Choose AgentKit if…
- You need full control over UX, auth, and hosting—embedding a production chat surface with ChatKit and orchestrating multi-step flows in Agent Builder.
- You want observability and reliability with OpenAI’s Evals to guard quality before you scale users.
Choose Responses API if…
- You’re integrating AI without a new surface, or you’re powering backend automations/services—with optional first-party tools like web search, file search, or computer use.
How they fit together (common architectures)
- ChatGPT app front-end + your agent back-end
- Build the UI & discovery in Apps SDK, but delegate heavy lifting to an AgentKit workflow via your API/MCP server. This lets you iterate agent logic in Agent Builder while the app remains simple and on-brand in ChatGPT.
- Embedded product chat + Responses automations
- Use ChatKit for your app’s interface; call the Responses API for model reasoning and tools (web/file/computer). Handy when you’re not targeting the ChatGPT directory but still want first-party capabilities.
- MCP-first integration strategy
- Expose your systems via MCP once; reuse across Apps SDK (ChatGPT app) and AgentKit (your product). This minimizes duplicate connector work.
Procurement-grade differences (what legal, IT, and product care about)
Distribution & governance
- Apps SDK: Must meet App Developer Guidelines; currently test in Developer Mode, with public submissions later this year. Good for reach (OpenAI cites 800M+ ChatGPT users).
- AgentKit / Responses: Your own compliance and product policies apply; no ChatGPT directory review.
UI ownership
- Apps SDK: Native UI rendered by ChatGPT per OpenAI design guidelines. Faster to ship in-chat, but the surface is curated.
- AgentKit: ChatKit gives you a production-ready chat component you theme and host.
- Responses: You build/own the UI entirely.
Quality assurance
- AgentKit: Evals are first-class for agent reliability.
- Apps SDK: Validate discovery and UI via Developer Mode before submission.
Decision tree (use it as a checklist)
Is ChatGPT the surface you want for users?
- Yes → Start with Apps SDK (then decide if your logic lives in MCP or calls out to your AgentKit backend).
- No → Build in your product with AgentKit or Responses API.
Do you need multi-step, agentic workflows with guardrails & evals?
- Yes → AgentKit (Agent Builder + Evals), optionally surfaced through ChatKit.
- No → A Responses integration or a simple Apps SDK app may be enough.
Is speed-to-market in ChatGPT critical (lead gen, support, guided commerce)?
- Yes → Apps SDK (preview today; submissions later this year).
- No → AgentKit/Responses give you maximum control over UX and deployment.
Example scoping (map work to platforms)
Apps SDK
- Deliverables: MCP tool contracts; in-chat UI flows; Developer Mode test plan; submission prep.
AgentKit
- Deliverables: Agent graph in Agent Builder; ChatKit embed; Evals suite; observability.
Responses API
- Deliverables: API orchestration; tool policies (web/file/computer); logging & fallbacks
Bottom line (how to choose fast)
- Need ChatGPT distribution + native in-chat UI? → Apps SDK (preview).
- Need production agents inside your app with evals & an embeddable UI? → AgentKit.
- Need programmable access for services/automations without a new UI? → Responses API.