Enact LLM Guide

Enact: Containerized tools with structured I/O for AI agents.

Commands

enact run ./tool --input "key=value"      # Run local tool
enact run ./tool --args '{"key":"value"}' # Run with JSON
enact run author/tool --input "x=y"       # Run installed tool
enact install author/tool                 # Install to project
enact install author/tool -g              # Install globally
enact search "query"                      # Find tools
enact sign ./tool && enact publish ./tool # Publish

Tool Structure

my-tool/
├── enact.md    # Required: YAML frontmatter + docs
└── main.py     # Your code (any language)

enact.md Template

---
enact: "2.0.0"
name: "namespace/category/tool-name"
version: "1.0.0"
description: "What it does"
from: "python:3.12-slim"
build: "pip install requests pandas"
command: "python /workspace/main.py ${input}"
timeout: "30s"

inputSchema:
  type: object
  properties:
    input:
      type: string
      description: "Input description"
  required: [input]

outputSchema:
  type: object
  properties:
    result:
      type: string

env:
  API_KEY:
    description: "API key"
    secret: true
  LOG_LEVEL:
    description: "Log level"
    default: "info"

tags: [category, keywords]
---

# Tool Name

Documentation here.

Field Reference

Field	Required	Description
name	Yes	namespace/category/tool
description	Yes	What it does
command	No*	Shell command with ${param} substitution
from	No	Docker image (default: alpine:latest)
build	No	Build commands (string or array), cached
inputSchema	No	JSON Schema for inputs
outputSchema	No	JSON Schema for outputs
env	No	Environment vars (secret: true for keyring)
timeout	No	Max runtime (default: 30s)
version	No	Semver version
tags	No	Discovery keywords

*Tools without command are LLM instruction tools (markdown interpreted by AI).

Examples by Language

Python

from: "python:3.12-slim"
build: "pip install pandas"
command: "python /workspace/main.py ${input}"

Node.js

from: "node:20-alpine"
build: "npm install"
command: "node /workspace/index.js ${input}"

Rust

from: "rust:1.83-slim"
build: "rustc /workspace/main.rs -o /workspace/app"
command: "/workspace/app ${input}"

Go

from: "golang:1.22-alpine"
build: "go build -o /workspace/app /workspace/main.go"
command: "/workspace/app ${input}"

Shell (no build)

command: "echo 'Hello ${name}'"

Source Code Pattern

Always output JSON matching outputSchema:

#!/usr/bin/env python3
import sys, json

input_val = sys.argv[1]
result = {"result": input_val.upper()}
print(json.dumps(result))

Secrets

env:
  API_KEY:
    description: "API key"
    secret: true  # Stored in OS keyring, not .env

User sets: enact env set API_KEY --secret --namespace myorg/tools

Access in code via environment variable: os.environ['API_KEY']

Two Tool Types

1. Container tools (has command): Runs in Docker, deterministic
2. Instruction tools (no command): Markdown body interpreted by LLM

Workflow

# 1. Create
mkdir my-tool && cd my-tool
# Create enact.md + source files

# 2. Test
enact run . --input "test=value"

# 3. Publish
enact auth login
enact sign .
enact publish .

Checklist

• [ ] name: namespace/category/tool format
• [ ] description: clear, searchable
• [ ] inputSchema: validates inputs
• [ ] outputSchema: documents output
• [ ] from: pinned image version (not latest)
• [ ] build: installs dependencies
• [ ] command: uses ${param} for inputs
• [ ] Source outputs valid JSON