// Docs //

From browse to trace

A plain-English guide to what each feature does and what to click first.

[ MARKETPLACE ][ BUILDER ][ SANDBOX ][ TRACES ]

What am I supposed to do?

quick start

Browse

Find a skill in Marketplace that matches the work you want done.

Inspect

Read the README, permissions, compatibility, versions, and eval score.

Run

Use Sandbox to test the skill with a safe prompt and real files.

Trace

Open the trace to see what the run touched, blocked, and produced.

Publish

Use Builder when you want to upload or author your own skill.

[ START HERE ]

Create a Skill

Marketplace

Choose a skill, inspect permissions, then run it before installing.

Builder

Upload a SKILL.md, .skill file, zip, or folder. Apply suggested fixes before publishing.

Sandbox

Run a skill with virtual provider mode or an approved real-shell sandbox command.

Evals

Save repeatable tests so each skill version can prove quality over time.

Traces

Review the evidence from each run: permissions, tools, terminal output, files, and artifacts.

Install

Export packages for Codex, Claude, VS Code, OpenCode, Grok, and Antigravity.

What is a skill?

[ CORE ]

Learn how this page works

Portable agent packages

open guide

A skill is a portable instruction package for an AI agent. It explains when to use the skill, what workflow to follow, what permissions are needed, and what examples prove it works.

Try this

Upload a folder that contains SKILL.md, README.md, references, scripts, and assets. Builder will parse it and suggest formatting fixes.

Why it matters

Clear skills make agent behavior repeatable. The marketplace adds trust by showing permissions, versions, evals, runs, traces, and install targets.

Read the page as

Left to right: input, controls, evidence, output. If something fails, check the evidence panel before changing the skill.

SKILL.md

The instruction file that tells the agent how to do a focused job.

Permissions

The safety gates that say what the skill may read, write, execute, or access.

Runs

A live test of the skill against a prompt and workspace files.

Traces

The record of the run, including events, output, artifacts, and blocked actions.