ChatGPT Apps + Apps SDK land with MCP, but early dev reports flag issues
OpenAI launched ChatGPT Apps with an Apps SDK built on the Model Context Protocol to bring third‑party services into ChatGPT, while developer reports highlight early stability, tooling, and tracing gaps to plan around. [This overview](http://www.monitorsaintpaul.com/premium/stacker/stories/chatgpt-apps-are-here-what-openairsquos-new-apps-sdk-means-for-marketers-and-developers,148140) details how ChatGPT is now a platform where users invoke partners like Booking.com, Canva, Coursera, Figma, Expedia, Spotify, and Zillow from chat, with an Apps SDK that extends MCP and a roadmap for an app directory, monetization, and enterprise-tier support later this year. Community threads surface implementation gotchas: guidance on [when to trigger your MCP](https://community.openai.com/t/how-to-tell-openai-when-to-use-your-mcp/1375779), missing tool-call visibility with the [Agents SDK Realtime model](https://community.openai.com/t/tool-calls-not-visible-in-traces-with-openai-agents-sdk-realtime-model/1375794), Agent Mode setup stalls on a [GitHub connector](https://community.openai.com/t/bug-agent-mode-stuck-on-setting-up-desktop-github-connector-never-starts/1375749), and a [Responses API shell tool availability bug](https://community.openai.com/t/shell-tool-in-responses-api-not-available-bug/1375776). Image workflows also need care: questions on [gpt-5.3-chat image rate limits with previous_response_id](https://community.openai.com/t/question-about-image-rate-limits-with-gpt-5-3-chat-when-using-previous-response-id/1375787) and intermittent 400s for `image_url` on [fine-tuned gpt-4o-2024-08-06](https://community.openai.com/t/intermittent-400-image-url-is-only-supported-by-certain-models-on-fine-tuned-gpt-4o-2024-08-06-was-working-previously/1375694); fine-tuning threads on [DPO validation loss plateaus](https://community.openai.com/t/why-is-my-dpo-validation-loss-not-decreasing/1375785) and [overfitting with GPT‑3.5 Turbo](https://community.openai.com/t/overfitting-issues-in-finetuning-gpt-3-5-turbo/710356) reinforce the need for tight evals and guardrails.