Which tool annotation identifies public web access in ChatGPT plugins?

Definition: openWorldHint: true is the tool annotation that identifies access to the public internet or open-ended external entities. OpenAI documents false for a tool limited to a bounded private account or workspace, even when that service is externally hosted.

Key facts

• OpenAI says ChatGPT and Codex use the metadata supplied for a tool when deciding when to call it, according to its metadata guidance.
• openWorldHint: true covers public internet access, open-ended external entities, and read-only tools such as web search.
• openWorldHint: false describes a tool limited to a bounded private account or workspace, including one hosted by an external service.
• OpenAI's plugin guidelines require openWorldHint to be an explicit Boolean in each tool's annotations object.
• The MCP tools specification describes annotations as optional properties about tool behavior and tells clients to treat them as untrusted unless they come from trusted servers.
• The same specification says a human should be able to deny tool invocations.

How do you set the annotation?

The annotation belongs inside the tool's annotations object. OpenAI's plugin guidelines require explicit Boolean values for readOnlyHint, destructiveHint, and openWorldHint on each tool. A public web search can therefore expose the relevant field as:

{
  "annotations": {
    "openWorldHint": true
  }
}

The value describes the tool's reach. It is separate from readOnlyHint, which OpenAI's metadata guidance assigns to retrieval or computation that does not create, update, delete, or send data outside the conversation.

What is the distinction between true and false?

Use true when the tool can reach a public or open-ended set of entities. OpenAI gives public internet access, arbitrary destinations, and read-only web search as examples. A search tool can be read-only and open-world at the same time because these annotations describe different properties.

Use false when the tool stays within a bounded private account, workspace, or catalog. The service can still be externally hosted. OpenAI's guidelines state that connectivity alone does not determine this value, so an external host is not by itself evidence for true.

What does the hint change in ChatGPT?

The metadata is part of the information ChatGPT and Codex use when deciding when to call a tool. MCP describes tools as model-controlled: a language model can locate and invoke them automatically from its understanding of the context and the user's prompts. The protocol also leaves implementations free to choose their interface pattern.

The hint is a description of intended tool behavior. OpenAI's plugin guidelines say annotations inform client safeguards, but do not grant permission or replace authentication, authorization, scope checks, input validation, or human confirmation for irreversible operations. The MCP specification likewise requires clients to treat annotations as untrusted unless they come from trusted servers.

How should you review a plugin tool?

Read the tool's actual scope alongside its annotation. A tool that searches the public web or accepts arbitrary public destinations fits the documented true case. A tool that reads one private workspace or account fits the documented false case, even if the API endpoint is on another company's infrastructure. Check readOnlyHint separately because a tool can be both read-only and open-world.

Related terms

• ChatGPT search
• Does ChatGPT always search the web before answering?
• Where can I see the sources behind a ChatGPT search answer?
• ChatGPT-User

FAQ

Should read-only web search use false?

No. OpenAI explicitly includes read-only tools such as web search in the openWorldHint: true case. Read-only describes changes to the environment, while open-world describes the entities the tool can access.

Should an externally hosted private workspace use true?

No. A bounded private account or workspace may use openWorldHint: false even when the service is externally hosted. The documented distinction is the tool's interaction scope.

Does openWorldHint grant public web access?

No. The annotation informs client safeguards. OpenAI's guidelines keep permission, authentication, authorization, scope checks, validation, and human confirmation as separate controls.

Sources

1. OpenAI, "Optimize Metadata", undated
2. OpenAI, "Plugin guidelines", undated
3. Model Context Protocol, "Tools", undated

---

Check how a page reads to answer engines. Open the GEO score analyzer