ChatGPT Apps Builder
Build documented ChatGPT apps that pair an MCP server with a widget UI.
The skill requires official-docs-first work, accurate tool-impact annotations, CSP handling, externalized secrets, idempotency, and explicit developer setup; no malicious or destructive default is evident. However, permission isolation, sensitive-data flow, rollback, and user-confirmation boundaries are incomplete, so 8 points are deducted.
The workflow, archetype taxonomy, validation ladder, and fallback paths are coherent. There is no committed test suite or reproducible execution evidence, core behavior depends on an external docs MCP service, and abnormal-input diagnostics remain limited; the static ceiling yields 8 points.
Audience, app types, trigger prompts, non-fit cases, and expected inputs are described fairly clearly. Chinese-language support and mainland-China reachability are not addressed, while core documentation access depends on developers.openai.com MCP, so 5 points are deducted.
The documentation has useful layering, progressive references, installation guidance, and stable terminology. The repository README says the repository is deprecated; the individual skill license is unknown, and versioning, changelog, maintenance ownership, and update path are unclear, so 7 points are deducted.
The claimed output—app classification, tool planning, scaffolding, validation reporting, and setup guidance—is directly useful and covers the main development stages. No representative output, runtime validation, or comparative-benefit evidence is included, so the static ceiling limits this to 7.
The skill and its reference files make the main rules auditable and distinguish validation levels. There is no committed test coverage, CI evidence, or third-party execution evidence, so verifiability is limited to 4.
- Core documentation retrieval depends on the external OpenAI docs MCP service; provide a local-docs or manual fallback for mainland-China reachability or service failures.
- The repository is marked deprecated, and the individual skill license, versioning, and maintenance ownership are unclear; verify the replacement repository and licensing before installation.
- The static materials do not prove that generated MCP servers, widgets, or the /mcp endpoint actually run.
What it does & when to use it
This skill supports designing, scaffolding, refactoring, and troubleshooting ChatGPT Apps SDK applications. It covers MCP servers, widget interfaces, the MCP Apps bridge, ChatGPT compatibility APIs, app metadata, content security policy, and domain settings. Its workflow requires consulting current OpenAI documentation and suitable official examples before generating code. It fits developers who need a tool plan, a working-project scaffold, or a local testing and connector setup path.
It classifies an app as tool-only, vanilla-widget, react-widget, interactive-decoupled, or submission-ready, then plans tool names, schemas, annotations, and outputs. It recommends an official example, an ext-apps example, or a local Node scaffold; generates MCP resource registration, tool handlers, metadata, and widget UI; explains MCP Apps bridge and window.openai integration, CSP configuration, local development, HTTPS tunneling, and ChatGPT Developer Mode setup; and validates the result against a minimum working repository contract with static, syntax, or runtime checks where feasible.
- A developer starting a ChatGPT app from scratch can use it to plan tools and generate an MCP server plus widget UI.
- A team adapting an existing Apps SDK example can use it to reconcile tool annotations, resource metadata, CSP, and URI versioning with current documentation.
- A developer choosing between React and vanilla HTML can use it to select an appropriate app archetype and starting example.
- A local developer connecting an MCP endpoint to ChatGPT Developer Mode can use it for HTTPS tunneling and refresh instructions.
- A team preparing a public ChatGPT app can use it to review hosting, submission, and pre-review requirements.
Pros & cons
- Covers the workflow from app classification and tool planning through server, widget, and validation.
- Explicitly prioritizes current OpenAI documentation and official examples.
- Combines the portable MCP Apps bridge with ChatGPT-specific window.openai compatibility features.
- Includes engineering checks for CSP, resource metadata, idempotency, and a minimum working repository contract.
- Its scope is specialized around ChatGPT Apps SDK, MCP, and related documentation.
- The supplied source does not provide a test suite, pinned versions, or a verified platform matrix.
- The repository README marks the repository as deprecated and points to the OpenAI Plugins repository for current examples.
- The supplied material does not establish the license for this individual skill directory.
How to install
The skill is located at skills/.curated/chatgpt-apps/ in openai/skills. The README documents installation of curated skills with $skill-installer chatgpt-apps in Codex; restart Codex after installation so it can load the skill. The same README states that the repository is deprecated and directs users to the OpenAI Plugins repository for current examples.
How to use
In Codex, invoke $openai-docs first, then pair it with this skill. Example: Use $chatgpt-apps with $openai-docs to scaffold a ChatGPT app for <use case> with a <TS/Python> MCP server and <React/vanilla> widget. You can also ask it to plan tools first and then generate the MCP server and widget. For local testing, the skill recommends running a local /mcp endpoint, exposing it through an HTTPS tunnel, and enabling Developer Mode under ChatGPT Settings → Apps & Connectors → Advanced settings.