New Feature Released: Custom Tools for Agents

[HighnessAtharva]

Custom tools are here. You can now add your existing bash, Python or Node.js scripts as tools your agent can call.

Open an agent, go to the Tools tab and click Create tool.

Create a Tool

Fill in the dialog:

1. Tool name: lowercase letters, numbers and underscores. Start with a letter.
2. Script file: a .sh, .py or .js file up to 64 KiB. The file extension selects the interpreter.
3. Description: tell the agent when to use the tool.

Then click Add input for each argument the agent must give your script. Each input has a name, a type, a description and a Required switch. The example above has one required Text input named city.

How a Script Gets Its Input

The agent sends the inputs to the script's stdin as JSON. With no inputs, stdin contains {}. The script parses the JSON itself.

For the city input, the script receives:

{ "city": "Mumbai" }

Whatever the script prints to stdout becomes the output the agent gets back.

Example Scripts

A bash version of the weather tool:

#!/usr/bin/env bash

city=$(jq -r '.city | @uri')
curl --fail --silent --show-error --max-time 20 "https://wttr.in/$city?format=3"

The Python version in the screenshot above reads stdin with json.load(sys.stdin) and prints the response.

Use the Tool

After you click Create tool, the tool appears in the Custom tools list with its description, script name and input count.

The agent now has the tool in its tool list. Ask for it in chat, for example "Check weather in Mumbai". The agent calls check_weather with city=Mumbai.

Inspect Every Call in Traces

Open Lens, then Traces, to see each tool call. The check_weather span shows the tool arguments and the tool result, so you can confirm what the agent sent and what your script returned.

Tell Us What You Think

Reply here with the scripts you turn into tools and anything that blocks you.

Some tenant details in the screenshots are redacted.