chat · framework · product

ChatGPT Canvas and Apps SDK

OpenAI dropped its model-written side panel (Canvas) from current models and bet on developer-built widgets in ChatGPT via the Apps SDK, now an MCP Apps host.

watchevidence: mixedby OpenAIdevelopers.openai.com/apps-sdk/reference ↗

ChatGPT Canvas and the Apps SDK

  • Maker: OpenAI
  • URL: https://developers.openai.com/apps-sdk/ ; https://help.openai.com/en/articles/9624314-model-release-notes
  • Status (10 October 2026):
    • Canvas launched 3 October 2024 (GPT-4o with Canvas). OpenAI’s model release notes say that from 28 May 2026 “canvas will no longer be available in GPT-5.5 Instant or GPT-5.5 Thinking”; writing and coding moved into chat as “writing blocks” and “code blocks”. It remains on legacy models until they are retired. No product-wide retirement has been announced.
    • Apps SDK launched in preview at DevDay in October 2025 (SEP-1865 says November 2025), built on MCP. ChatGPT became “fully compatible” with MCP Apps on 22 February 2026; window.openai is now described as a compatibility layer plus ChatGPT-specific extras. Apps can be distributed as Codex plugins (March 2026). Closed host, open SDK docs.
    • ChatKit (embeddable chat UI) continues; Agent Builder is being retired in November 2026 (from the first landscape pass; not re-verified here).

What it is

Two different answers from one lab. Canvas was a side panel where the model wrote and edited a document or code with the person, including running simple HTML/React previews. The Apps SDK lets developers ship apps that appear inside ChatGPT: an MCP server whose tools come with HTML widgets rendered inline, picture-in-picture or fullscreen.

The problem it’s solving

Canvas: chat is a bad place to co-edit a long object. Apps SDK: ChatGPT as a platform, where third-party services show real interfaces in the conversation instead of text descriptions of them.

Its path / bet

OpenAI has stepped back from the “model writes a page beside the chat” path (Canvas dropped from the current models in favour of inline blocks) and doubled down on the developer-authored widget path (Apps SDK, now converged on MCP Apps). That’s the opposite of Anthropic’s 2026 direction, where model-written pages (Claude Code Artifacts) grew. The disagreement between the two labs is one of the clearest signals in this cluster.

How it works

  • An app is an MCP server. Tools declare a UI template (_meta.ui.resourceUri, alias openai/outputTemplate) and CSP.
  • ChatGPT renders the widget in a sandboxed iframe and supplies toolInput, toolOutput (structuredContent; “the model reads it as-is”, so keep it concise) and toolResponseMetadata (hidden from the model).
  • The widget keeps widgetState via setWidgetState, calls tools with callTool, posts into the chat with sendFollowUpMessage, requests display modes, uploads and selects files.
  • Users choose how often connected apps ask permission (June 2026).

Strengths

  • The largest consumer chat surface; distribution through an app directory.
  • Converged on the open standard, so the same server works in Claude and VS Code.
  • Clear control over what the model sees (structuredContent) versus what only the widget sees (_meta), which is the same “data that skips the model” idea fictty has, applied to tool results.

Weaknesses / limits

  • Widgets are code a developer ships, reviewed ahead of time. Not a medium for an agent’s throwaway screen.
  • Chat-only surface; nothing for terminal agents (Codex CLI gets plugins, not widgets).
  • Canvas’s withdrawal from current models suggests OpenAI found model-written side panels less valuable than inline blocks, or too costly to maintain; OpenAI hasn’t said why.

Relation to fictty

Watch. No direct overlap with a terminal runtime. The Canvas reversal is a data point for the blog post: one lab is betting the model should write pages, the other pulled back to inline structured blocks and developer-built apps.

Could fictty adopt it instead of building?

No. It’s a host feature of ChatGPT.

What fictty should take from it

  • Inline blocks beat side panels for most answers. OpenAI’s retreat to writing and code blocks says the heavy surface should be summoned, not default. fictty screens should be opt-in and short-lived, and an agent should be told when not to open one.
  • Separate model-visible and UI-only data in the format. The structuredContent/_meta split is a clean pattern; fictty’s get --state could offer a compact model view versus the full state.
  • Display modes (inline, PiP, fullscreen) map neatly onto a pane, a callout layer and a full screen; worth borrowing the names.

Sources

Not verified: whether o3 (the last legacy model with Canvas) was in fact retired on 26 August 2026, and whether the Apps SDK is still labelled preview.