Tool Development

​

Creating a Tool

​

The Tool Editor

​

Keyboard Shortcuts

​

Writing Tools in TypeScript

import { z } from "zod";

// Define the input schema
export const input = z.object({
  name: z.string().describe("The user's name"),
  greeting: z.string().optional().describe("Custom greeting prefix"),
});

// Define the output schema
export const output = z.object({
  message: z.string().describe("The greeting message"),
});

// Implement the tool logic
export async function run(data: z.infer<typeof input>) {
  const prefix = data.greeting ?? "Hello";
  return { message: `${prefix}, ${data.name}!` };
}

• Import z from "zod" — it is pre-installed in the sandbox.
• Export input and output as Zod schemas. These are used for validation and to generate JSON Schema for the API.
• Export run as the entry point. It receives the validated input and must return an object matching the output schema.
• You can import any packages listed in the ToolSet’s Packages tab.
• Environment variables are accessible via process.env.

​

Writing Tools in Python

from pydantic import BaseModel, Field

class Input(BaseModel):
    name: str = Field(description="The user's name")
    greeting: str = Field(default="Hello", description="Custom greeting prefix")

class Output(BaseModel):
    message: str = Field(description="The greeting message")

def run(input: Input) -> Output:
    return Output(message=f"{input.greeting}, {input.name}!")

• Define Input and Output as Pydantic BaseModel subclasses.
• Define run as the entry point. It receives the validated input and must return an instance of the Output model.
• You can import any packages listed in the ToolSet’s Packages tab.
• Environment variables are accessible via os.environ.

​

Input and Output Schemas

1. Validation — The API validates the input (required fields and property types) against the input schema before execution. The output is not validated.
2. Documentation — Schemas are exposed through the API and MCP endpoints so that AI agents understand what a tool expects and returns.

​

Auto-Generate Metadata

• Name — a human-readable name inferred from the code.
• Slug — derived from the code.
• Description — a summary of what the tool does.

​

AI Copilot

Create a tool that posts a message to a Slack channel
Add a limit parameter to the input
Fix the error

• Generate or change code — The agent writes or edits the implementation. Changes it makes through its editing tools appear in the editor as they happen; for code shown in the chat, click Apply to put it into the editor, which also saves the tool. Copy copies the snippet instead.
• Fix a failed test — The latest run result (status, input, output, error, stdout and stderr) is sent with every message. After a failed test, an Auto Fix Error button appears that asks the agent to fix the error based on that result.
• Run and fill in test input — The agent can run the tool itself and can fill the Input tab of the Test panel with suggested test input, typically after a successful run.
• Conversations — Conversations are saved per tool and per user. The most recent one loads when the panel opens; the header has a conversation selector, New conversation and Export conversation.

​

API Panel

• The Organization Endpoint (POST .../v1/orgs/{orgId}/toolsets/{slug}/tools/{toolSlug}/run) with a copyable curl example whose sample input is generated from the tool’s input schema.
• The Public Endpoint (POST .../v1/public/{orgId}/{slug}/run/{toolSlug}) with its curl example when the ToolSet is public, or a note that the ToolSet is private.
• A reminder that each execution consumes 1 credit.

​

Saving and Deleting

• Save — Click the Save button or press Cmd/Ctrl + S. Changes are saved to the draft state and do not affect published versions.
• Delete — Click the Delete button in the header. A confirmation dialog will appear. Deleting a tool removes it from the draft state. If it was included in published versions, those snapshots remain unchanged.

​

Best Practices

• Keep tools focused. Each tool should do one thing well. Use the ToolSet to group related tools.
• Write descriptive schemas. Good field descriptions help AI agents call your tools correctly.
• Test before publishing. Use the Test panel to verify behavior with various inputs before publishing a new version.
• Use environment variables for secrets. Never hard-code API keys or tokens in tool code. Store them in the ToolSet’s Environment tab.
• Leverage the Copilot for iteration. After a failed test run, the Copilot has access to the error output and can suggest targeted fixes.