ADL CLI

A command-line interface for generating enterprise-ready A2A (Agent-to-Agent) servers from Agent Definition Language (ADL) files.

⚠️ Early Development Warning: This project is in its early stages of development. Breaking changes are expected and acceptable until we reach a stable version. Use with caution in production environments.

Table of Contents

• Overview
• Installation
  • npm / npx (Recommended)
  • Install Script
  • From Source
  • Using Go Install
  • Pre-built Binaries
• Quick Start
• Usage
  • Commands
  • Init Command
  • Generate Command
• Agent Definition Language (ADL)
• Generated Project Structure
• Sandbox Environments
• Enterprise Features
• Artifacts Support
• GitHub Issue Templates
• Examples
• Template System & Architecture
• Customizing Generation with .adl-ignore
• Configurable Acronyms
• Post-Generation Hooks
• Development
• Roadmap
• License
• Support

Overview

The ADL CLI helps you build enterprise-ready A2A agents quickly by generating complete project scaffolding from YAML-based Agent Definition Language (ADL) files. It eliminates boilerplate code and ensures consistent patterns across your agent implementations.

Key Features

• 🚀 Rapid Development - Generate complete projects in seconds
• 📋 Schema-Driven - Use YAML Agent Definition Language files (ADL) to define your agents
• 🎯 Enterprise Ready - Single unified template with AI integration and enterprise features
• 🔐 Enterprise Features - Authentication, SCM integration, and audit logging
• 🛠️ Smart Ignore - Protect your implementations with .adl-ignore files
• ✅ Validation - Built-in ADL schema validation
• 🛠️ Interactive Setup - Guided project initialization with extensive CLI options
• 🔗 Structured Services - Type-safe dependency injection with interfaces and factory functions
• ⚙️ Configuration Management - Automatic environment variable mapping with proper naming conventions
• 🔧 CI/CD Generation - Automatic GitHub Actions workflows with semantic-release CD pipelines
• 🏗️ Sandbox Environments - Flox and DevContainer support for isolated development
• 🎣 Post-Generation Hooks - Customize build, format, and test commands after generation
• 🤖 Multi-Provider AI - OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek, Cohere, Cloudflare, Moonshot, Ollama, Ollama Cloud, and Nvidia support
• 📁 Artifacts Support - Integrated filesystem and MinIO object storage for artifact management

Installation

npm / npx (Recommended)

Most developers already have Node.js - run adl without a prior install via @inference-gateway/adl-cli, a thin wrapper that downloads the matching native binary on first use:

npx @inference-gateway/adl-cli init my-agent
npx @inference-gateway/adl-cli generate --file agent.yaml --output ./agent
npx @inference-gateway/adl-cli validate agent.yaml

Or install it globally:

npm install -g @inference-gateway/adl-cli
adl --help

Prebuilt binaries are available for Linux and macOS on x64/arm64. The npm package version tracks the native binary version one-to-one.

Install Script

Use our install script to automatically download and install the latest binary:

curl -fsSL https://raw.githubusercontent.com/inference-gateway/adl-cli/main/install.sh | bash

Or download and run the script manually:

wget https://raw.githubusercontent.com/inference-gateway/adl-cli/main/install.sh
chmod +x install.sh
./install.sh

Install Options:

• Install specific version: ./install.sh --version v1.0.0
• Custom install directory: INSTALL_DIR=~/bin ./install.sh
• Show help: ./install.sh --help

From Source

git clone https://github.com/inference-gateway/adl-cli.git
cd adl-cli
go install .

Using Go Install

go install github.com/inference-gateway/adl-cli@latest

Nix Flake

Run the latest version directly without installing:

nix run github:inference-gateway/adl-cli

Or pin a specific version:

nix run github:inference-gateway/adl-cli/v0.41.3

Build and add it to your profile:

nix profile install github:inference-gateway/adl-cli/v0.41.3

Enter a development shell with go, go-task, golangci-lint, gopls, and goreleaser available:

nix develop github:inference-gateway/adl-cli

Flox

Pin adl to a specific version inside a Flox environment by adding it to your .flox/env/manifest.toml:

[install]
adl.flake = "github:inference-gateway/adl-cli/v0.41.3"

Then activate the environment:

flox activate

Pre-built Binaries

Download pre-built binaries from the releases page.

Quick Start

1. Initialize a New Project

# Interactive project setup - creates ADL manifest
adl init my-weather-agent

# Generate project code from the manifest
adl generate --file agent.yaml --output ./test-my-agent

2. Implement Your Business Logic

The generated project includes TODO placeholders for your implementations:

// TODO: Implement weather API logic
func GetWeatherTool(ctx context.Context, args map[string]any) (string, error) {
    city := args["city"].(string)
    // TODO: Replace with actual weather API call
    return fmt.Sprintf(`{"city": "%s", "temp": "22°C"}`, city), nil
}

3. Build and Run

cd test-weather-agent
task build
task run

Usage

Commands

| Command | Description | | --------------------- | ------------------------------------------------------------------ | | adl init [name] | Create ADL manifest file interactively with options | | adl generate | Generate project code from ADL file with CI/CD and sandbox support | | adl validate [file] | Validate an ADL file against the complete schema |

Init Command

The adl init command provides a interactive wizard for creating ADL manifest files:

# Interactive ADL manifest creation
adl init my-weather-agent

# Use defaults for all prompts
adl init my-agent --defaults

# Non-interactive with specific configuration
adl init my-agent \
  --name "Weather Agent" \
  --description "Provides weather information" \
  --provider deepseek \
  --model deepseek-v4-flash \
  --language go \
  --flox

Init Command Options

The init command supports extensive configuration options:

Project Settings:

• --defaults - Use default values for all prompts
• --path - Project directory path
• --name - Agent name
• --description - Agent description
• --version - Agent version

Agent Configuration:

• --type - Agent type (ai-powered/minimal)
• --provider - AI provider (openai/anthropic/google/groq/mistral/deepseek/cohere/cloudflare/moonshot/ollama/ollama_cloud/nvidia/minimax)
• --model - AI model name
• --system-prompt - System prompt for the agent
• --max-tokens - Maximum tokens (integer)
• --temperature - Temperature (0.0-2.0)

Capabilities:

• --streaming - Enable streaming responses
• --notifications - Enable push notifications
• --history - Enable state transition history

Server Configuration:

• --port - Server port (integer)
• --debug - Enable debug mode

Language-Specific Options:

• --language - Programming language (go/rust/typescript)

Go Options:

• --go-module - Go module path (e.g., github.com/user/project)
• --go-version - Go version (e.g., 1.26.2)

Rust Options:

• --rust-package-name - Rust package name
• --rust-version - Rust version (e.g., 1.94)
• --rust-edition - Rust edition (e.g., 2024)

TypeScript Options:

• --typescript-name - TypeScript package name

Environment Options:

• --flox - Enable Flox environment
• --devcontainer - Enable DevContainer environment

Pipeline / AI Options (declarative, written into the manifest as false by default):

• --ai - Shortcut for the init wizard: writes spec.development.ai.orchestrators.claudecode.enabled: true into the generated agent.yaml. Every other per-agent toggle (codex, gemini, opencode, infer) stays off; edit agent.yaml after init to enable additional agents (see Per-agent AI assistants).
• --ci - Sets spec.scm.ci: true (generate CI workflow on adl generate)
• --cd - Sets spec.scm.cd: true (generate CD pipeline + semantic-release on adl generate)

Generate Command

# Generate project from ADL file
adl generate --file agent.yaml --output ./test-my-agent

# Overwrite existing files (respects .adl-ignore)
adl generate --file agent.yaml --output ./test-my-agent --overwrite

# Generate with CI workflow configuration
adl generate --file agent.yaml --output ./test-my-agent --ci

# Generate with CloudRun deployment configuration
adl generate --file agent.yaml --output ./test-my-agent --deployment cloudrun

# Generate with CloudRun deployment and CD pipeline
adl generate --file agent.yaml --output ./test-my-agent --deployment cloudrun --cd

Generate Flags

| Flag | Description | | ------------------ | ---------------------------------------------------------------------------------- | | --file, -f | ADL file to generate from (default: "agent.yaml") | | --output, -o | Output directory for generated code (default: ".") | | --template, -t | Template to use (default: "minimal") | | --overwrite | Overwrite existing files (respects .adl-ignore) | | --ci | Generate CI workflow configuration (GitHub Actions). Overrides spec.scm.ci. | | --cd | Generate CD pipeline configuration with semantic-release. Overrides spec.scm.cd. | | --deployment | Generate deployment configuration (kubernetes, cloudrun) |

Declarative equivalents: --ci and --cd are mirrored by spec.scm.ci and spec.scm.cd. The CLI flag is OR'd on top of the manifest value (passing the flag wins; omitting it falls back to the manifest). AI assistants are entirely manifest-driven via the per-agent toggles in spec.development.ai.orchestrators

• see the matrix below. adl init writes all toggles as false by default - they're opt-in. Generated files (CLAUDE.md, GEMINI.md, AGENTS.md, .github/workflows/ci.yml, .github/workflows/cd.yml, .github/workflows/claude.yml, .github/workflows/codex.yml, .github/workflows/gemini.yml, .releaserc.yaml) are tagged linguist-generated=true in .gitattributes so they collapse in pull request diffs.

CI Generation Features:

• Automatic Provider Detection: Detects GitHub from ADL spec.scm.provider (GitLab support planned)
• Language-Specific Workflows: Tailored CI configurations for Go, Rust, and TypeScript
• Version Integration: Uses language versions from ADL configuration
• Task Integration: Leverages generated Taskfile for consistent build processes
• Caching: Includes service caching for faster builds

CD Generation Features:

• Semantic Release Integration: Automatic versioning based on conventional commits
• Multi-Language Support: Builds and tests for Go, Rust, and TypeScript projects
• Container Publishing: Builds and pushes Docker images to GitHub Container Registry
• Manual Dispatch: CD workflow triggered manually via GitHub Actions
• Changelog Generation: Automatic CHANGELOG.md generation with release notes
• GitHub Releases: Creates GitHub releases with appropriate tagging
• Deployment Integration: Supports automatic deployment to Kubernetes and Cloud Run after successful releases

AI Integration Features:

The ADL CLI honours the per-agent toggles under spec.development.ai.orchestrators (the per-agent toggles were introduced in ADL schema v0.8.0 and moved under orchestrators in adl#27). Each entry is independent and defaults to false:

spec:
  development:
    ai:
      orchestrators:
        claudecode:
          enabled: true # generates CLAUDE.md + .github/workflows/claude.yml
        codex:
          enabled: false # would generate AGENTS.md + .github/workflows/codex.yml
        gemini:
          enabled: false # would generate GEMINI.md + .github/workflows/gemini.yml
        opencode:
          enabled: false # would generate AGENTS.md (no upstream action yet)
        infer:
          enabled: false # would generate AGENTS.md (no upstream action yet)

Per-agent AI assistants

| Agent toggle | Docs file the agent reads | GitHub Actions workflow generated? | | ------------ | ------------------------- | --------------------------------------------------------------------------------- | | claudecode | CLAUDE.md | yes (.github/workflows/claude.yml, uses anthropics/claude-code-action) | | codex | AGENTS.md (shared) | yes (.github/workflows/codex.yml, uses openai/codex-action) | | gemini | GEMINI.md | yes (.github/workflows/gemini.yml, uses google-github-actions/run-gemini-cli) | | opencode | AGENTS.md (shared) | no upstream action yet - docs only | | infer | AGENTS.md (shared) | no workflow scaffolded yet - docs only |

• AGENTS.md is generated once and is shared by every enabled agent that reads from it (codex, opencode, infer); the file's contents are agent-agnostic.
• CLAUDE.md and GEMINI.md are agent-specific and only appear when the matching toggle is on.
• If no toggles are enabled, no AI docs or workflows are emitted.
• Pre-v0.8.0 manifests using spec.development.ai.enabled: true, and the flat per-agent shape spec.development.ai.<agent> (pre-orchestrators), are no longer accepted - adl validate and adl generate fail with a migration hint pointing at spec.development.ai.orchestrators. Move the toggle to the specific agent you want (e.g. orchestrators.claudecode.enabled: true).
• When claudecode is enabled, sandbox environments (Flox, DevContainer) also gain the claude-code CLI / extension automatically.

Deployment Generation Features: The --deployment flag generates platform-specific deployment configurations:

• CloudRun Deployment: Creates a deploy task in the root Taskfile.yml for gcloud deployment
  • Supports both Google Container Registry (GCR) and GitHub Container Registry (GHCR)
  • Configurable resources (CPU, memory), scaling (min/max instances), and service options
  • Uses direct gcloud commands for truly serverless deployment (no Kubernetes required)
  • Automatic container building with Docker or Cloud Build integration
• Kubernetes Deployment: Creates k8s/deployment.yaml with standard Kubernetes manifests
  • Enterprise-ready configurations with resource limits and health checks
  • ConfigMap and Secret integration for environment variables
  • Service and Ingress configurations for load balancing

Agent Definition Language (ADL)

ADL files use YAML to define your agent's configuration, capabilities, and tools.

The canonical schema lives in the inference-gateway/adl repository - that repo is the single source of truth for the ADL specification. This CLI vendors a pinned copy at internal/schema/schema.json (refresh with task fetch-schema).

Example ADL File

apiVersion: adl.inference-gateway.com/v1
kind: Agent
metadata:
  name: weather-agent
  description: "Provides weather information for cities worldwide"
  version: "1.0.0"
spec:
  capabilities:
    streaming: true
    pushNotifications: false
    stateTransitionHistory: false
  agent:
    provider: "" # Choose: openai, anthropic, google, groq, mistral, deepseek, cohere, cloudflare, moonshot, ollama, ollama_cloud, nvidia, minimax
    model: "" # Specify default model name for chosen provider
    systemPrompt: "You are a helpful weather assistant."
    maxTokens: 4096
    temperature: 0.7
  tools:
    - name: get_weather
      description: "Get current weather for a city"
      schema:
        type: object
        properties:
          city:
            type: string
            description: "City name"
          country:
            type: string
            description: "Country code"
        required:
          - city
  server:
    port: 8080
    debug: false
  language:
    go:
      module: "github.com/example/weather-agent"
      version: "1.26.2"
  acronyms: # Optional: Custom acronyms for better code generation
    - api
    - json
    - xml

ADL Schema

The complete ADL schema includes:

• metadata: Agent name, description, and version
• capabilities: Streaming, notifications, state history
• config: Structured configuration sections with environment variable mapping
• services: Service services with interfaces, factories, and type definitions
• agent: AI provider configuration (OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek, Cohere, Cloudflare, Moonshot, Ollama, Ollama Cloud, Nvidia)
• tools: Function-call definitions with JSON schemas, validation, and service injection support
• skills: Markdown playbooks (id + optional bare, version, source) pulled from the skills registry, fetched as a full directory from a GitHub repo (shorthand or URL), or scaffolded locally; advertised on the agent card and prepended to the system prompt at runtime
• server: HTTP server configuration with authentication support
• language: Programming language-specific settings (Go, Rust, TypeScript) and configurable acronyms
• scm: Source control management configuration (GitHub, GitLab)
• sandbox: Development environment configuration (Flox, DevContainer)
• deployment: Platform-specific deployment configuration (Kubernetes, Cloud Run)

Complete ADL Example

apiVersion: adl.inference-gateway.com/v1
kind: Agent
metadata:
  name: advanced-agent
  description: "Enterprise agent with full feature set"
  version: "1.0.0"
spec:
  capabilities:
    streaming: true
    pushNotifications: true
    stateTransitionHistory: true
  agent:
    provider: deepseek
    model: deepseek-v4-flash
    systemPrompt: |
      You are a helpful assistant with enterprise capabilities.
      Always prioritize security and compliance.
    maxTokens: 8192
    temperature: 0.3
  config:
    database:
      connectionString: "postgresql://user:pass@localhost:5432/db"
      maxConnections: "10"
      timeout: "30s"
    notifications:
      slackWebhook: "https://hooks.slack.com/services/..."
      emailApiKey: "your-email-api-key"
      retryAttempts: "3"
  services:
    database:
      type: service
      interface: DatabaseService
      factory: NewDatabaseService
      description: PostgreSQL database service for persistent storage
    notifications:
      type: service
      interface: NotificationService
      factory: NewNotificationService
      description: Multi-channel notification service
  tools:
    - name: query_database
      description: "Execute database queries with validation"
      inject:
        - logger
        - database
      schema:
        type: object
        properties:
          query:
            type: string
            description: "SQL query to execute"
          table:
            type: string
            description: "Target table name"
          limit:
            type: integer
            description: "Result limit"
            maximum: 1000
        required:
          - query
          - table
    - name: send_notification
      description: "Send multi-channel notifications"
      inject:
        - logger
        - notifications
      schema:
        type: object
        properties:
          recipient:
            type: string
            description: "Recipient identifier"
          message:
            type: string
            description: "Message content"
          priority:
            type: string
            enum:
              - low
              - medium
              - high
              - critical
          channel:
            type: string
            enum:
              - email
              - slack
              - teams
              - webhook
        required:
          - recipient
          - message
          - priority
          - channel
  server:
    port: 8443
    debug: false
    auth:
      enabled: true
  language:
    go:
      module: "github.com/company/advanced-agent"
      version: "1.26.2"
  scm:
    provider: github
    url: "https://github.com/company/advanced-agent"
  deployment:
    type: cloudrun
    cloudrun:
      image:
        registry: gcr.io
        repository: advanced-agent
        tag: latest
        useCloudBuild: true
      resources:
        cpu: "2"
        memory: 1Gi
      scaling:
        minInstances: 1
        maxInstances: 100
        concurrency: 1000
      service:
        timeout: 3600
        allowUnauthenticated: false
        serviceAccount: advanced-agent@PROJECT_ID.iam.gserviceaccount.com
        executionEnvironment: gen2
      environment:
        LOG_LEVEL: info
        ENVIRONMENT: production
  development:
    sandbox:
      flox:
        enabled: true

Extra dependencies (spec.language.<lang>.vendor)

Every language config block accepts an optional vendor section that lets the manifest extend the generator's built-in dependency set. Use deps for runtime/production dependencies and devdeps for development-only ones. The exact meaning of devdeps depends on the language - see the mapping table below. Each entry must be <package>@<version> using the target language's native package and version syntax - the schema validates the shape up front (^\S+@\S+$) and points at the offending key if you mistype it (e.g. spec.language.go.vendor.deps.0).

Conflict policy: generator built-ins always win. If your manifest lists a package that the generator already pins (e.g. github.com/inference-gateway/adk for Go, tokio for Rust), the vendor entry is silently dropped and a ⚠️ vendor … collides with built-in … warning is printed to stderr. This prevents accidental downgrades of the core runtime SDK.

Output mapping per language:

| Language | deps lands in | devdeps lands in | | ---------- | ----------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Go | go.mod require block | go.mod tool directive (executable dev tools: code generators, linters, etc.) plus an // indirect entry in require so the module is downloadable. Test libraries that you import (testify, go-cmp, …) belong in deps, not devdeps. | | Rust | Cargo.toml [dependencies] | Cargo.toml [dev-dependencies] | | TypeScript | package.json dependencies | package.json devDependencies |

For Go, supply the full tool package path (the binary's main package, e.g. golang.org/x/tools/cmd/stringer) with a version. After generation, run go mod tidy so Go normalises the indirect require entry to the actual module root.

Examples per language:

# Go: uuid for runtime, stringer + mockgen as dev tools.
spec:
  language:
    go:
      module: github.com/example/agent
      version: "1.26.2"
      vendor:
        deps:
          - github.com/google/[email protected]
          - github.com/stretchr/[email protected] # imported by *_test.go
        devdeps:
          - golang.org/x/tools/cmd/[email protected]
          - github.com/golang/mock/[email protected]

# Rust: regex at runtime, mockall + pretty_assertions for tests.
spec:
  language:
    rust:
      packageName: agent
      version: "1.94.1"
      edition: "2024"
      vendor:
        deps:
          - [email protected]
        devdeps:
          - [email protected]
          - [email protected]

# TypeScript: axios at runtime, vitest + @types/node for tests.
spec:
  language:
    typescript:
      packageName: "@example/agent"
      nodeVersion: "20"
      vendor:
        deps:
          - [email protected]
        devdeps:
          - "@types/[email protected]"
          - [email protected]

Extra sandbox dependencies (spec.development.deps)

spec.development.deps is the cross-cutting equivalent of the per-language vendor.deps block above: it lets you install tools into the development sandbox (Flox, devcontainer) that don't belong to any single language's package manager. Use it for things like deno, kubectl, terraform, awscli, or any other CLI you want available inside the dev shell. Each entry follows the <package>@<version> shape, validated up front by the schema (^\S+@\S+$).

spec:
  development:
    sandbox:
      flox:
        enabled: true
      devcontainer:
        enabled: true
    deps:
      - deno@^2.7.14
      - kubectl@^1.36.1
      - terraform@^1.15.3

Merge semantics: additive. The per-language toolchain that each sandbox template already installs (go/cargo/nodejs, plus git, docker, go-task) is always emitted; spec.development.deps entries are appended on top, sorted alphabetically so the generated file diffs stay stable on re-run. Duplicate package names inside deps are deduped (first occurrence wins on version). When a deps entry's package name collides with one of the template's built-ins (e.g. [email protected]), the generator prints a ⚠️ spec.development.deps … collides with a Flox built-in … warning to stderr but still renders the user entry - the maintainer's pin wins on version conflict.

Output mapping per backend:

| Backend | Generated file | Per-entry rendering | | --------------- | ------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | flox | .flox/env/manifest.toml | A pair of TOML lines under [install]: <pkg>.pkg-path = "<pkg>" / <pkg>.version = "<version>". Resolved against Nixpkgs at activation time. | | devcontainer | .devcontainer/devcontainer.json | Added to the features block as ghcr.io/devcontainers-extra/features/apt-packages:1 with the comma-joined <pkg>=<version> list. Resolution is best-effort against apt. | | dockerCompose | (out of scope for this release; see issue #154) | n/a |

If a package isn't published under its short attribute path in Nixpkgs, or isn't available as an apt package on the devcontainer base image, add the file to your project's .adl-ignore and hand-author the override - this is the same escape hatch we use for any sandbox file the template can't infer.

Worked example per backend:

# Flox: pin deno + kubectl + terraform alongside the Go toolchain
spec:
  language:
    go: { module: github.com/example/agent, version: "1.26.2" }
  development:
    sandbox:
      flox:
        enabled: true
    deps:
      - [email protected]
      - [email protected]
      - [email protected]

# Devcontainer: same deps, rendered as an apt-packages feature
spec:
  language:
    go: { module: github.com/example/agent, version: "1.26.2" }
  development:
    sandbox:
      devcontainer:
        enabled: true
    deps:
      - [email protected]
      - [email protected]

Skills vs. Tools

The ADL spec distinguishes two complementary concepts:

• Tools (spec.tools) are function-call entrypoints with explicit JSON schemas. They are generated as code in the target language and registered with the agent's toolbox. The model invokes them by name with structured arguments.
• Skills (spec.skills) are markdown playbooks (with YAML frontmatter) that describe when and how to use the tools. Each is written to its own directory at skills/<id>/SKILL.md in the generated project, advertised on the agent card so orchestrators can discover them, and prepended to the system prompt at runtime. The directory layout matches Anthropic's agent skills convention - bare skills can ship arbitrary scripts, templates, or reference material alongside SKILL.md.

A skill entry is small:

spec:
  skills:
    - id: data-analysis # pulled from registry.inference-gateway.com/skills/
      version: 0.1.0 # optional pin
    - id: report-writing # pulled at the default version
    - id: company-policy # scaffolded locally (not fetched)
      bare: true
      name: company-policy
      description: "Internal compliance rules to follow"
      license: Proprietary # optional SPDX id or "Proprietary"
      tags: [policy, compliance]
    - id: pdf # pull a full skill directory from GitHub
      source: anthropics/skills/pdf
    - id: skill-creator
      source: [email protected] # pin to a tag/branch/sha

Resolution rules

• bare: true → the CLI scaffolds skills/<id>/SKILL.md with frontmatter from the manifest and a TODO body that you author by hand. The whole skills/<id>/ directory is listed in .adl-ignore, so any bundled scripts, templates, or resources you drop alongside SKILL.md are preserved on regeneration.
• source: set → the source must resolve to a public GitHub directory (a /tree/<ref>/<path> URL, or one of the shorthand forms below). The CLI pulls the entire directory - SKILL.md, reference docs, bundled scripts, anything else - and writes it to skills/<id>/. Non-github.com URLs are rejected so the same code path always produces a complete skill bundle, not a stray markdown file.
• Otherwise → fetch https://registry.inference-gateway.com/skills/<id>[/<version>].md (becomes skills/<id>/SKILL.md). Override the registry with ADL_SKILLS_REGISTRY. Registry-by-id currently ships SKILL.md only; if you need bundled assets, use source: to point at a GitHub directory.

Licensing

license is optional on every skill entry. When set, it must be one of the SPDX identifiers enumerated in the schema (MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause, GPL-2.0, GPL-3.0, LGPL-2.1, LGPL-3.0, MPL-2.0, ISC, CC0-1.0, CC-BY-4.0, CC-BY-SA-4.0, Unlicense) or the literal string Proprietary for closed-source skills. The resolver mirrors the value into the generated SKILL.md frontmatter so the licence travels with the playbook - shipping a separate LICENSE file alongside SKILL.md is optional. When the ADL entry and the fetched frontmatter both set license, the value in the ADL manifest wins.

Use adl generate --offline to skip network access - every non-bare skill must already be cached at ~/.adl/skills-cache/<id>@<ref>/ (where <ref> is the pinned tag/branch, or latest for an unpinned registry fetch).

source: shorthand grammar

Every form below resolves to a GitHub tree/<ref>/<path> URL. An optional @<tag> suffix pins a branch, tag, or commit SHA; omit it to use the default main branch.

| Shorthand | Expands to | | --------------------------------- | ----------------------------------------------------------------------- | | <skill> | https://github.com/inference-gateway/skills/tree/main/skills/<skill> | | <skill>@<tag> | https://github.com/inference-gateway/skills/tree/<tag>/skills/<skill> | | <owner>/<repo>/<skill> | https://github.com/<owner>/<repo>/tree/main/skills/<skill> | | <owner>/<repo>/<skill>@<tag> | https://github.com/<owner>/<repo>/tree/<tag>/skills/<skill> | | Full https://github.com/... URL | passed through unchanged |

Concrete examples:

# Default inference-gateway/skills, latest main:
- id: skill-creator
  source: skill-creator

# Default inference-gateway/skills, pinned to a tag:
- id: skill-creator
  source: [email protected]

# Different repo (Anthropic's official skill library):
- id: pdf
  source: anthropics/skills/pdf

# Different repo, pinned to a commit SHA:
- id: pdf
  source: anthropics/skills/pdf@abc1234

# Full URL for anything that doesn't fit the shorthand:
- id: custom
  source: https://github.com/my-org/my-repo/tree/release/path/to/skill

The 3-segment form assumes a skills/<id>/ subdirectory inside the repo (the convention used by both inference-gateway/skills and anthropics/skills). If your repo lays skills out differently, pass the full URL.

Runtime: AVAILABLE SKILLS manifest + on-demand Read

The generated agent advertises skills to the LLM via a frontmatter-only manifest, not by inlining SKILL.md bodies. At startup it walks first-level subdirectories under skills/ (overridable with A2A_SKILLS_DIR), parses each <id>/SKILL.md's YAML frontmatter, and appends an AVAILABLE SKILLS: block to the system prompt:

AVAILABLE SKILLS:
Skills are reusable instructions for specific tasks. When a task matches a
skill's description, read the SKILL.md file at the listed path using the Read
tool, then follow its instructions.

- incident-response: Use this when the user reports a production incident...
  Path: skills/incident-response/SKILL.md
- pdf: Fill in PDF forms and extract structured data from PDFs.
  Path: skills/pdf/SKILL.md

The model loads each SKILL.md body on demand via the Read built-in tool, and executes any bundled scripts via Bash / Write / Edit. A skills-using agent must therefore list - id: read in spec.tools and set spec.config.tools.read.enabled: true - the validator enforces this; see Reserved built-in tools.

Reserved built-in tools

spec.tools accepts five reserved IDs that map to framework-supplied implementations:

| Reserved ID | Generated as | Purpose | | ----------- | --------------------- | ------------------------------------------------------------------------------------- | | read | tools/read.go etc. | Read a file (file_path, optional offset/limit). | | bash | tools/bash.go etc. | Execute a shell command (subject to whitelist + timeout). | | write | tools/write.go etc. | Write content to a file (creates parent dirs). | | edit | tools/edit.go etc. | Replace a unique string in a file (old_string → new_string). | | fetch | tools/fetch.go etc. | Fetch an http(s) URL (whitelist, max-bytes cap, optional save-to-disk inside /tmp). |

Opt in by listing the id alone - the generator owns name, description, and the JSON schema:

spec:
  tools:
    - id: read
    - id: bash
    - id: query_database # user tool: full entry still required
      name: query_database
      description: "..."
      schema: { type: object, ... }

All five built-ins default to enabled: false. Activate them via the reserved namespace spec.config.tools.<id>:

spec:
  config:
    tools:
      read:
        enabled: true
        max_lines: 2000 # offset/limit default window
        allowed_roots: [] # empty = project-wide
      bash:
        enabled: true
        whitelist: [ls, cat, grep, jq]
        timeout_seconds: 30
      write:
        enabled: false # listed but explicitly disabled
      edit:
        enabled: true
      fetch:
        enabled: true
        allowed_domains: # whitelist of hosts (empty = unrestricted, discouraged)
          - example.com
          - .api.dev # entries starting with "." match any subdomain
        max_bytes: 10485760 # 10 MiB cap on response body (default)
        timeout_seconds: 30 # total request timeout (default)
        download_dir: /tmp # root for save_path writes (default /tmp)
        allow_downloads: false # set true to allow writing response bodies to disk

Values are baked into the generated constructor as compile-time literals - there's no ToolsConfig struct in config/config.go because reserved-namespace sections are intentionally skipped. The validator decodes each spec.config.tools.<id> block into the built-in's typed shape and rejects unknown keys (typos like tymeout_seconds fail with spec.config.tools.bash.tymeout_seconds).

Runtime overrides for Bash (read inside tools/bash.go):

• A2A_BASH_DISABLED=1 is a kill switch - overrides enabled: true back to false.
• A2A_BASH_WHITELIST=ls,cat,grep overrides the compile-time whitelist.

Runtime overrides for Fetch (resolution precedence: env > compile-time literal > default-disabled):

• Go: TOOLS_FETCH_ENABLED, TOOLS_FETCH_ALLOWED_DOMAINS, TOOLS_FETCH_MAX_BYTES, TOOLS_FETCH_TIMEOUT_SECONDS, TOOLS_FETCH_DOWNLOAD_DIR, TOOLS_FETCH_ALLOW_DOWNLOADS (envconfig-style; comma-separated for lists).
• Rust: A2A_FETCH_DISABLED=1 (kill switch), A2A_FETCH_ALLOWED_DOMAINS, A2A_FETCH_MAX_BYTES, A2A_FETCH_TIMEOUT_SECONDS, A2A_FETCH_DOWNLOAD_DIR, A2A_FETCH_ALLOW_DOWNLOADS.

The Fetch tool supports GET and HEAD only. Optional save_path writes the response body to a path resolved under download_dir - absolute paths and parent-directory traversal (..) are rejected, and the request fails unless allow_downloads: true. Bodies (and on-disk files) are capped at max_bytes; oversized responses are truncated and the result payload sets "truncated": true. The Go template uses only the standard library (net/http); the Rust template adds reqwest (rustls-tls + json features) to Cargo.toml automatically when - id: fetch is present in spec.tools.

Resolution precedence at runtime: env > compile-time literal > built-in default (disabled).

Service Injection & Configuration Management

The ADL CLI provides a sophisticated service injection system with structured configuration management. This system improves testability, separation of concerns, and provides type-safe configuration with environment variable mapping.

Structured Service System

Define services with explicit types, interfaces, and factory functions. The system supports both built-in services (like logger) and custom service services:

spec:
  config:
    googleCalendar:
      scopes: "https://www.googleapis.com/auth/calendar"
      credentialsPath: "/secrets/credentials.json"
    cache:
      ttl: "3600"
      maxEntries: "1000"
  services:
    googleCalendar:
      type: service
      interface: CalendarService
      factory: NewCalendarService
      description: Google Calendar API service for managing calendar events
    cache:
      type: service
      interface: CacheRepository
      factory: NewCacheRepository
      description: High-performance caching layer for API responses
  tools:
    - name: create_event
      description: "Create a new calendar event"
      inject:
        - logger # Built-in, always available
        - googleCalendar # Custom service
        - cache # Custom service
      schema:
        type: object
        properties:
          title:
            type: string
            description: "Event title"
          start:
            type: string
            description: "Start time (ISO 8601)"
        required: [title, start]

Configuration Management

The configuration system generates type-safe structs with automatic environment variable mapping:

Generated Configuration (config/config.go):

type Config struct {
    // Core application settings
    Environment string `env:"ENVIRONMENT"`

    // A2A configuration
    A2A serverConfig.Config `env:",prefix=A2A_"`

    // Custom configuration sections
    Cache          CacheConfig          `env:",prefix=CACHE_"`
    GoogleCalendar GoogleCalendarConfig `env:",prefix=GOOGLE_CALENDAR_"`
}

type GoogleCalendarConfig struct {
    CredentialsPath string `env:"CREDENTIALS_PATH"`
    Scopes          string `env:"SCOPES"`
}

type CacheConfig struct {
    MaxEntries string `env:"MAX_ENTRIES"`
    Ttl        string `env:"TTL"`
}

Environment Variables:

• GOOGLE_CALENDAR_CREDENTIALS_PATH="/secrets/google-creds.json"
• GOOGLE_CALENDAR_SCOPES="https://www.googleapis.com/auth/calendar"
• CACHE_MAX_ENTRIES="1000"
• CACHE_TTL="3600"

Config Subsection Injection

In addition to injecting entire configuration objects, you can inject specific config subsections directly into skills using dotted notation. This provides type-safe access to focused configuration scopes.

Example ADL Configuration:

spec:
  config:
    database:
      connectionString: "postgresql://localhost:5432/db"
      maxConnections: "10"
      timeout: "30s"
    email:
      apiKey: ""
      fromAddress: "[email protected]"
      provider: "sendgrid"
  services:
    database:
      type: service
      interface: DatabaseService
      factory: NewDatabaseService
      description: PostgreSQL database service
  tools:
    - name: export_report
      description: "Export data and email report"
      inject:
        - logger
        - database
        - config.email # Inject only the email config subsection
      schema:
        type: object
        properties:
          recipient:
            type: string
        required: [recipient]

Generated Skill Code:

type ExportReportSkill struct {
    logger   *zap.Logger
    database database.DatabaseService
    email    *config.EmailConfig  // Type-safe access to email config only
}

func NewExportReportSkill(
    logger *zap.Logger,
    database database.DatabaseService,
    email *config.EmailConfig,
) server.Tool {
    skill := &ExportReportSkill{
        logger:   logger,
        database: database,
        email:    email,
    }
    // ...
}

func (s *ExportReportSkill) ExportReportHandler(ctx context.Context, args map[string]any) (string, error) {
    // Direct access to email config subsection
    apiKey := s.email.APIKey
    fromAddress := s.email.FromAddress
    provider := s.email.Provider

    // ... implementation
}

Main Registration:

// In main.go - config subsection is passed directly
exportReportSkill := skills.NewExportReportSkill(l, databaseSvc, &cfg.Email)
toolBox.AddTool(exportReportSkill)

Benefits of Config Subsection Injection:

• Scoped Access: Skills only receive the configuration they need, following principle of least privilege
• Type Safety: Compile-time validation ensures config fields exist
• Clear Dependencies: Explicit declaration of which config sections each skill requires
• Easier Testing: Mock specific config subsections without full config object
• Better Separation: Skills don't have access to unrelated configuration
• Auto-Validation: ADL CLI validates that injected config sections exist in spec.config

Injection Patterns:

inject:
  - logger # Built-in logger service
  - config # Entire config object (*config.Config)
  - config.database # Database config subsection (*config.DatabaseConfig)
  - config.email # Email config subsection (*config.EmailConfig)
  - myService # Custom service from spec.services

Service Architecture

The service injection system generates:

1. Built-in Logger: Automatically available as *zap.Logger without declaration
2. Type-Safe Configuration: Structured config with environment variable mapping
3. Service Interfaces: Custom service packages with interface definitions
4. Factory Functions: Constructor functions that receive logger and configuration
5. Automatic Registration: Services are automatically wired into skills
6. File Protection: Generated service files are automatically added to .adl-ignore

Generated Structure

my-agent/
├── config/
│   └── config.go                    # Type-safe configuration with env mapping
├── internal/
│   ├── logger/
│   │   └── logger.go               # Built-in logger factory
│   ├── googleCalendar/
│   │   └── googleCalendar.go       # Calendar service with interface
│   └── cache/
│       └── cache.go                # Cache service with interface
├── tools/
│   ├── create_event.go             # Function-call tools with injected services
│   └── list_events.go
├── skills/
│   ├── calendar-workflow/          # Markdown playbooks loaded into the system prompt
│   │   └── SKILL.md
│   └── meeting-summary/
│       └── SKILL.md
└── .adl-ignore                     # Protects custom implementations

Generated Service Code

Each service generates a package with interface and factory:

Example internal/googleCalendar/googleCalendar.go:

type CalendarService interface {
    // TODO: Define your CalendarService interface methods
    CreateEvent(ctx context.Context, event *Event) error
    ListEvents(ctx context.Context, query *Query) ([]*Event, error)
}

type calendarService struct {
    logger *zap.Logger
    config *config.Config
}

func NewCalendarService(logger *zap.Logger, cfg *config.Config) (CalendarService, error) {
    // TODO: Implement CalendarService initialization
    return &calendarService{
        logger: logger,
        config: cfg,
    }, nil
}

Skill Integration

Skills automatically receive injected services as constructor parameters:

Example skills/create_event.go:

type CreateEventSkill struct {
    logger    *zap.Logger
    calendar  googleCalendar.CalendarService
    cache     cache.CacheRepository
}

func NewCreateEventSkill(logger *zap.Logger, calendar googleCalendar.CalendarService, cache cache.CacheRepository) *CreateEventSkill {
    return &CreateEventSkill{
        logger:   logger,
        calendar: calendar,
        cache:    cache,
    }
}

Benefits

• Type Safety: Structured configuration with compile-time validation
• Environment Variables: Automatic mapping with proper naming conventions
• Interface-Based Design: Testable services with clear contracts
• Separation of Concerns: Configuration separate from service definitions
• Language Agnostic: Works across Go, Rust, and TypeScript
• Hot Reload: Configuration changes via environment variables
• Security: No secrets in code, environment-based configuration
• Scalability: Easy to add new services and configuration sections

Best Practices

1. Configuration: Use environment variables for secrets and environment-specific values
2. Interfaces: Define clear interfaces for testability and modularity
3. Factory Functions: Initialize services with proper error handling
4. Logging: Use the injected logger for consistent log formatting
5. Testing: Create mock implementations of service interfaces
6. Documentation: Document interface methods and configuration options

Generated Project Structure

The ADL CLI generates project scaffolding tailored to your chosen language:

Go Project Structure

my-go-agent/
├── main.go                    # Main server setup
├── go.mod                     # Go module definition
├── config/
│   └── config.go              # Centralized application configuration
├── internal/
│   └── logger/
│       └── logger.go          # Built-in logger factory
├── tools/                     # Function-call tool implementations
│   ├── query_database.go      # Individual tool files (TODO placeholders)
│   └── send_notification.go
├── skills/                    # Skill directories (SKILL.md + optional bundled assets)
│   ├── incident-response/     # Loaded into the system prompt at startup
│   │   └── SKILL.md
│   └── support-handoff/
│       └── SKILL.md
├── Taskfile.yml               # Development tasks (build, test, lint)
├── Dockerfile                 # Container configuration
├── .adl-ignore                # Files to protect from regeneration
├── .well-known/
│   └── agent-card.json        # Agent capabilities (auto-generated)
├── .github/                   # GitHub-specific configurations
│   ├── workflows/             # Generated when using --ci flag
│   │   ├── ci.yml             # GitHub Actions CI workflow
│   │   └── cd.yml             # GitHub Actions CD workflow (with --cd flag)
│   ├── dependabot.yml         # Generated when scm.dependabot: true
│   └── ISSUE_TEMPLATE/        # Generated when issue_templates: true
│       ├── bug_report.md      # Bug report template
│       ├── feature_request.md # Feature request template
│       └── refactor_request.md # Refactoring request template
├── .releaserc.yaml            # Semantic-release configuration (with --cd flag)
├── k8s/
│   └── deployment.yaml        # Kubernetes deployment manifest
├── cloudrun/
│   └── deploy.sh              # CloudRun deployment script (with --deployment cloudrun)
├── .flox/                     # Generated when sandbox: flox
│   ├── env/manifest.toml
│   ├── env.json
│   ├── .gitignore
│   └── .gitattributes
├── .gitignore                 # Standard Git ignore patterns
├── .gitattributes             # Git attributes configuration
├── .editorconfig              # Editor configuration
├── CLAUDE.md                  # AI assistant instructions (spec.development.ai.orchestrators.claudecode.enabled: true)
└── README.md                  # Project documentation with setup instructions

Rust Project Structure

my-rust-agent/
├── src/
│   ├── main.rs                # Main application entry point
│   └── tools/                 # Function-call tool implementations
│       ├── mod.rs             # Module declarations
│       ├── query_database.rs  # Individual tool implementations
│       └── send_notification.rs
├── skills/                    # Skill directories (SKILL.md + optional bundled assets)
│   ├── incident-response/
│   │   └── SKILL.md
│   └── support-handoff/
│       └── SKILL.md
├── Cargo.toml                 # Rust package configuration
├── Taskfile.yml               # Development tasks
├── Dockerfile                 # Rust-optimized container
├── .adl-ignore                # Protection configuration
├── .well-known/
│   └── agent-card.json        # Agent capabilities
├── .github/workflows/         # CI configuration (with --ci)
│   ├── ci.yml                 # Rust-specific CI workflow
│   └── cd.yml                 # GitHub Actions CD workflow (with --cd flag)
├── .releaserc.yaml            # Semantic-release configuration (with --cd flag)
├── k8s/
│   └── deployment.yaml        # Kubernetes deployment
├── cloudrun/
│   └── deploy.sh              # CloudRun deployment script (with --deployment cloudrun)
├── CLAUDE.md                  # AI assistant instructions (spec.development.ai.orchestrators.claudecode.enabled: true)
└── README.md                  # Documentation

TypeScript Project Structure

my-typescript-agent/
├── src/
│   ├── index.ts               # Main server setup + background task worker
│   ├── config.ts              # Typed configuration loader (env-mapped)
│   ├── logger.ts              # Built-in logger factory
│   ├── services/              # Injectable service implementations
│   │   └── database.ts
│   └── tools/                 # Function-call tool implementations
│       ├── index.ts           # Toolbox wiring
│       ├── query_database.ts  # Individual tool implementations (TODO placeholders)
│       └── send_notification.ts
├── skills/                    # Skill directories (SKILL.md + optional bundled assets)
│   ├── incident-response/
│   │   └── SKILL.md
│   └── support-handoff/
│       └── SKILL.md
├── package.json               # Node package definition + scripts
├── tsconfig.json              # TypeScript compiler configuration
├── Taskfile.yml               # Development tasks
├── Dockerfile                 # Node-optimized container
├── .adl-ignore                # Protection configuration
├── .well-known/
│   └── agent-card.json        # Agent capabilities
├── .github/workflows/         # CI configuration (with --ci)
│   ├── ci.yml                 # TypeScript-specific CI workflow
│   └── cd.yml                 # GitHub Actions CD workflow (with --cd flag)
├── .releaserc.yaml            # Semantic-release configuration (with --cd flag)
├── k8s/
│   └── deployment.yaml        # Kubernetes deployment (with --deployment kubernetes)
├── CLAUDE.md                  # AI assistant instructions (spec.development.ai.orchestrators.claudecode.enabled: true)
└── README.md                  # Documentation

Universal Generated Files

All projects include these essential files regardless of language:

• .well-known/agent-card.json - A2A agent discovery and capabilities manifest
• Taskfile.yml - Unified task runner configuration for build, test, lint, run
• Dockerfile - Language-optimized container configuration
• k8s/deployment.yaml - Kubernetes deployment manifest
• deploy task in Taskfile.yml - CloudRun deployment task (when using --deployment cloudrun)
• .adl-ignore - Protects user implementations from overwrite
• CI Workflows - When using --ci flag, generates GitHub Actions workflows:
  • GitHub Actions: .github/workflows/ci.yml
  • GitLab CI: .gitlab-ci.yml (planned, not yet implemented)
• CD Workflows - When using --cd flag, generates continuous deployment:
  • GitHub Actions: .github/workflows/cd.yml
  • Semantic Release: .releaserc.yaml
• Development Environment - Based on sandbox configuration:
  • Flox: .flox/ directory with environment configuration when sandbox.flox.enabled: true
  • DevContainer: .devcontainer/devcontainer.json when sandbox.devcontainer.enabled: true
• AI Assistant Instructions - Per-agent toggles under spec.development.ai.orchestrators (see Per-agent AI assistants):
  • CLAUDE.md when spec.development.ai.orchestrators.claudecode.enabled: true
  • GEMINI.md when spec.development.ai.orchestrators.gemini.enabled: true
  • AGENTS.md (shared) when any of codex, opencode, or infer is enabled

CI Integration

When using the --ci flag, the ADL CLI generates GitHub Actions workflows for your project:

# Generate project with CI workflow
adl generate --file agent.yaml --output ./test-my-agent --ci

This creates a GitHub Actions workflow (.github/workflows/ci.yml) that includes:

• Automated Testing: Runs all tests on every push and pull request
• Code Quality: Format checking and linting
• Multi-Environment: Supports main and develop branches
• Caching: Go module caching for faster builds
• Task Integration: Uses the generated Taskfile for consistent build steps

The generated workflow automatically detects your Go version from the ADL file and configures the appropriate environment.

CD Integration

The ADL CLI can generate continuous deployment (CD) pipelines with semantic release automation:

# Generate project with CD pipeline
adl generate --file agent.yaml --output ./test-my-agent --cd

This creates a complete CD setup including:

• .releaserc.yaml - Semantic-release configuration with conventional commits
• .github/workflows/cd.yml - GitHub Actions CD workflow with manual dispatch

The generated CD pipeline includes:

• Semantic Versioning: Automatic version bumping based on conventional commit messages
• Release Automation: Creates GitHub releases with generated release notes
• Container Publishing: Builds and publishes Docker images to GitHub Container Registry
• Multi-Platform Builds: Supports both AMD64 and ARM64 architectures
• Language Detection: Automatically configures build steps based on your project language
• Change Detection: Only publishes releases when there are changes to release

CD Workflow Features

Manual Trigger: The CD workflow uses workflow_dispatch for controlled releases:

# Trigger via GitHub CLI
gh workflow run cd.yml

# Or trigger via GitHub Actions UI

Conventional Commits Support: The pipeline recognizes these commit types for versioning:

• feat: - Minor version bump (new features)
• fix: - Patch version bump (bug fixes)
• refactor:, perf:, ci:, docs:, style:, test:, build:, chore: - Patch version bump

Container Registry: Published images are available at:

ghcr.io/your-org/your-agent:latest
ghcr.io/your-org/your-agent:v1.0.0
ghcr.io/your-org/your-agent:1.0

CloudRun Deployment

The ADL CLI provides native support for deploying A2A agents to Google Cloud Run, offering a truly serverless deployment experience without Kubernetes complexity.

CloudRun Configuration

Configure CloudRun deployment in your ADL file:

spec:
  deployment:
    type: cloudrun
    cloudrun:
      image:
        registry: gcr.io # gcr.io or ghcr.io
        repository: my-agent # Repository name
        tag: latest # Image tag
        useCloudBuild: true # Use Cloud Build or local Docker
      resources:
        cpu: "2" # CPU allocation (0.1 to 8)
        memory: 1Gi # Memory limit (128Mi to 32Gi)
      scaling:
        minInstances: 0 # Minimum instances (0 to 1000)
        maxInstances: 100 # Maximum instances (1 to 1000)
        concurrency: 1000 # Max concurrent requests per instance
      service:
        timeout: 3600 # Request timeout in seconds
        allowUnauthenticated: true # Allow public access
        serviceAccount: my-agent@PROJECT_ID.iam.gserviceaccount.com
        executionEnvironment: gen2 # gen1 or gen2
      environment: # Custom environment variables
        LOG_LEVEL: info
        ENVIRONMENT: production

Container Registry Options

Google Container Registry (GCR):

image:
  registry: gcr.io
  repository: my-project/my-agent
  useCloudBuild: true # Automatically build and push

GitHub Container Registry (GHCR):

image:
  registry: ghcr.io
  repository: myorg/my-agent
  useCloudBuild: false # Skip Cloud Build, use pre-built image

Generated Deployment Script

When using --deployment cloudrun, the ADL CLI generates a deploy task in the Taskfile.yml that:

• Validates Environment: Checks for required PROJECT_ID and REGION variables
• Container Building: Uses Docker locally or Cloud Build based on configuration
• Direct gcloud Deployment: Uses gcloud run deploy for serverless deployment
• Configuration Summary: Displays all deployment settings for verification

CloudRun Deployment Workflow

# 1. Generate project with CloudRun deployment
adl generate --file agent.yaml --output ./my-agent --deployment cloudrun

# 2. Set required environment variables
export PROJECT_ID="my-gcp-project"
export REGION="us-central1"

# 3. Deploy to CloudRun
cd my-agent
task deploy

CloudRun with CI/CD

Generate CloudRun deployment with continuous deployment:

adl generate --file agent.yaml --deployment cloudrun --cd

This creates:

• CD Workflow: Automatically deploys to CloudRun after releases
• Environment Integration: Uses GitHub secrets for GCP authentication
• Multi-Environment Support: Deploy to different regions/projects

Required GitHub Secrets:

• GCP_SA_KEY: Service account key JSON
• GCP_PROJECT_ID: Google Cloud project ID
• GCP_REGION: Deployment region (e.g., us-central1)

CloudRun Benefits

• Truly Serverless: No Kubernetes clusters or infrastructure management
• Auto-Scaling: Scale to zero when idle, scale up automatically under load
• Pay-per-Use: Only pay for actual request processing time
• Global Edge: Deploy to multiple regions with traffic management
• Integrated Monitoring: Built-in logging, metrics, and tracing
• Custom Domains: HTTPS support with automatic SSL certificates

Example ADL Files

The CLI includes CloudRun example files:

# Validate CloudRun examples
adl validate examples/cloudrun-agent.yaml
adl validate examples/cloudrun-ghcr-agent.yaml

# Generate CloudRun projects
adl generate --file examples/cloudrun-agent.yaml --output ./cloudrun-test
adl generate --file examples/cloudrun-ghcr-agent.yaml --output ./ghcr-test

Cloudflare Workers Deployment

Cloudflare Workers run on the V8-isolate edge runtime and, like Vercel and unlike Kubernetes/CloudRun, deploy from source via wrangler rather than a prebuilt container image. This target is TypeScript-only - Workers execute JS/TS on the edge, so a Go or Rust agent does not produce Worker artifacts.

Cloudflare Configuration

spec:
  language:
    typescript:
      packageName: "my-agent"
      nodeVersion: "24"
  deployment:
    type: cloudflare
    cloudflare:
      name: my-agent # Worker (script) name
      accountId: "${CLOUDFLARE_ACCOUNT_ID}" # prefer a ${VAR} placeholder
      compatibilityDate: "2025-01-01" # defaults when omitted
      compatibilityFlags:
        - nodejs_compat # defaults to [nodejs_compat] when omitted
      routes:
        - agent.example.com/* # custom routes/domains
      workersDev: false # expose on *.workers.dev (omit to keep wrangler's default)
      environment: # plain-text wrangler [vars]
        LOG_LEVEL: info

Generated Artifacts

• wrangler.toml - rendered from the block (name, main, compatibility_date, compatibility_flags, account_id, workers_dev, routes, [vars]).
• src/worker.ts - a module-format Worker entrypoint scaffold exporting a fetch handler. It is added to .adl-ignore, so your completed handler survives subsequent adl generate runs.

compatibilityDate defaults to a stable value when omitted, and compatibilityFlags defaults to nodejs_compat (the scaffold relies on Node.js API compatibility).

Secrets

Secrets are never written to wrangler.toml. Use ${VAR} placeholders only in the environment block (plain-text [vars]), and set real secrets out-of-band:

wrangler secret put LLM_API_KEY

Deploy

# One-time setup (not installed by the scaffold)
pnpm add -D wrangler @cloudflare/workers-types
# then add "@cloudflare/workers-types" to compilerOptions.types in tsconfig.json

#