create-client-tool

SKILL.md

Create a Client Tool

Scaffold a new AtlasTool named $ARGUMENTS and wire it into the app.

Prerequisite

integrate-atlas-chat must already be complete: the app should vend the atlas-agent sources under src/atlas-agent/ (including react.ts) and have @sinclair/typebox installed as in that skill.

Background

Client tools let the Atlas Agent invoke logic that runs in the browser — rendering charts,
querying local state, showing UI panels, triggering navigation, etc. The agent decides when
to call the tool; the app executes it and returns a result.
The flow is:

Agent responds with a clientTool action

The library validates the arguments against the TypeBox schema

execute() runs in the browser and returns { output, details }

output (string) is sent back to the agent as the tool result

details (any shape) is available on message.toolCalls for the UI to render

Step 1 — Understand the codebase

Before writing anything, read:

The file where useAtlasChat is called (often src/App.tsx or a chat hook) to find where tools is passed — imports are typically from ./atlas-agent/react after integrate-atlas-chat

Any existing tool definitions to match the file/naming conventions

Step 2 — Define the tool

Create the tool as a typed constant. Use Type from @sinclair/typebox to define the parameters schema — this gives both compile-time types and runtime validation (same stack as the vendored atlas-agent from integrate-atlas-chat).

import { Type } from "@sinclair/typebox";
import type { AtlasTool } from "./atlas-agent/types";

export const myTool: AtlasTool = {
  name: "my_tool",            // snake_case — this is what the agent uses to invoke it
  description:
    "One sentence describing what this tool does and when the agent should call it.",
  parameters: Type.Object({
    exampleParam: Type.String({ description: "What this param is for" }),
    optionalNum: Type.Optional(Type.Number({ description: "..." })),
  }),
  execute: async (args) => {
    // args is fully typed from the schema above
    // Do the work here — call APIs, update state, render UI, etc.
    return {
      output: "Plain text summary sent back to the agent",
      details: {
        // Any structured data you want available in the UI via message.toolCalls
      },
    };
  },
};

Adjust the ./atlas-agent/... path if the tool file is not directly under src/ next to the atlas-agent folder (for example ../atlas-agent/types from src/tools/).

TypeBox quick reference

Schema
Usage

Type.String()
string

Type.Number()
number

Type.Boolean()
boolean

Type.Literal("foo")
exact value

Type.Union([Type.Literal("a"), Type.Literal("b")])
enum

Type.Array(Type.String())
string[]

Type.Object({ ... })
object

Type.Optional(...)
mark any field optional

Always add a description to each field — the agent uses these to understand what to pass.

Step 3 — Wire into useAtlasChat

Find the useAtlasChat call and add the tool to the tools array:

const { messages, send, ... } = useAtlasChat({
  client: isLoading ? null : sdk,
  agentExternalId: AGENT_EXTERNAL_ID,
  tools: [myTool],   // add here
});

Step 4 — Render tool results (if needed)

If the tool returns structured details, render them in the message list.
message.toolCalls is a ToolCall[] — one entry per tool call (client-side and server-side) in call order.

{msg.toolCalls?.map((tc, i) => (
  // tc.name    — tool name
  // tc.output  — the string sent back to the agent
  // tc.details — your structured data (cast to your known shape)
  <MyToolOutput key={i} data={tc.details as MyToolDetails} />
))}

Done

The agent can now invoke $ARGUMENTS. Describe what it does clearly in the description
field — the agent relies on that string to decide when and how to call the tool.