Skip to main content
Buttons gives agents deterministic escape hatches. Instead of hoping an LLM follows a multi-step procedure correctly, agents press buttons — self-contained, composable actions with clear inputs, outputs, and status. Each button wraps a single action: a shell script, an HTTP API call, a Python/Node snippet, or an agent instruction. Create it once with typed arguments and a timeout, press it whenever you need it, get structured JSON output every time.

​
Get started

Install

curl, Docker, Go, or Homebrew.

Quick start

Create and press your first button in 30 seconds.

​
Core Concepts

Buttons

The core unit of reusable work.

Drawers

Typed workflows that compose buttons.

Registry

Install and publish shared buttons, pinned per-button by version and content hash.

History

Local audit history for creates, edits, installs, and updates.

​
Button guides

Code

Inline shell, Python, or Node scripts.

HTTP API

GET/POST with URL templates and typed args.

File import

Import an existing script file.

Prompt

Attach an instruction for the consuming agent.

​
Workflows

Drawers chain buttons into typed workflows with refs, loops, sub-drawers, waits, and run history.

Workflows

Build multi-step automations from reusable buttons.

​
Triggers

Triggers run buttons and drawers automatically — on a schedule (cron), on local file changes (watch), or from an inbound webhook. Button triggers fire under buttons serve; drawer webhook triggers fire under buttons webhook listen.

Trigger concept

How buttons and drawers fire automatically.

Trigger guide

Commands and management model.

Hook

Run on local events — design proposal.

​
Operations

REST API server

Expose local buttons over HTTP with the same press contract.

​
CLI reference

Every command is auto-generated from the Cobra command tree. Flags, examples, and descriptions stay in sync with the binary.

CLI reference

Full command reference: create, press, list, history, delete, version, and more.

​
For deployed agents

Buttons ships as a static binary and a multi-arch Docker image. Bake it into your agent container with one line:
Your agent calls buttons press <name> --json as a subprocess and gets structured output. No SDK, no library import, no runtime dependency.