create-client-tool

A scaffolding skill for creating an AtlasTool, a client-side tool that an Atlas agent can invoke, and wiring it into the application. It is meant to be used whenever a new client tool is needed instead of hand-writing AtlasTool definitions or manually wiring them into useAtlasChat, and it handles the TypeBox schema, the execute function, and the hook wiring. It applies to tools that fetch data, render UI, call APIs, show charts, query local state, or perform any other browser-side action, and it triggers on phrases like AtlasTool, client tool, add tool, create tool, new tool, tool definition, and agent tool. The prerequisite is that the integrate-atlas-chat skill has already vendored the atlas-agent sources under src/atlas-agent/ and installed @sinclair/typebox.

Client tools let the Atlas Agent run logic in the browser while the agent decides when to call them. The documented flow is: the agent responds with a clientTool action, the library validates the arguments against the TypeBox schema, execute() runs in the browser and returns an object with output and details, the output string is sent back to the agent as the tool result, and the details (any shape) become available on message.toolCalls for the UI to render. The skill first reads the codebase to find where useAtlasChat is called and to match existing tool conventions.

The tool itself is defined as a typed AtlasTool constant whose parameters schema is built with Type from @sinclair/typebox, giving both compile-time types and runtime validation. A quick reference covers common TypeBox schemas including String, Number, Boolean, Literal, Union for enums, Array, Object, and Optional, and every field should carry a description because the agent relies on those to decide what to pass. The tool name is snake_case since that is what the agent uses to invoke it. After definition, the tool is added to the tools array of the useAtlasChat call, and if it returns structured details those can be rendered by mapping over message.toolCalls, where each ToolCall exposes name, output, and details. Clear description fields are emphasized throughout, since the agent depends on them to decide when and how to call the tool.