Raycast AI Extensions

AI extensions let Raycast AI call your extension's tools to answer questions and take actions ("@your-extension create an issue titled …"). You provide the tools; the model decides when to call them. Build on raycast-core (manifest + command model); this spoke covers the AI surface. Confirm current signatures at developers.raycast.com/ai.

Mental model

• A tool is a typed function the AI may invoke: it receives validated input and returns data the model uses to continue. Tools are the AI analog of commands.
• Declare tools in package.json → tools[] (each with a name, title, and description the model reads to decide relevance), with a matching implementation file that default-exports the tool function.
• The extension's description and any AI instructions steer when/how the AI uses your tools — write them for the model, not the user.

Designing good tools

• One clear job per tool. Prefer getIssues + createIssue over a mega-tool with a mode arg — the model routes better to focused tools.
• Strong descriptions. The tool/param descriptions are the model's only routing signal; say what it does, when to use it, and what each argument means.
• Typed inputs. Validate every argument; never trust the model's input blindly (same boundary discipline as user input).
• Read vs write. Mark/structure side-effecting tools so they require confirmation before mutating (create/delete/send). Read tools can run freely.