Agentfile is workflow-as-files for AI agents.
Current stable release: v0.1.0
Development continues on main branch.
Agentfile is an open specification for describing AI-powered workflows as plain files — agents, skills, steps, and rules — that any IDE agent (Cursor, Windsurf, GitHub Copilot, Claude Code, Cline, Roo) can read and execute.
Most AI workflows are locked inside proprietary tools, chat UIs, or framework-specific code. Agentfile makes workflows portable, versionable, and agent-agnostic — define once, run anywhere.
| Without Agentfile | With Agentfile |
|---|---|
| Workflow logic lives in the tool | Workflow logic lives in your repo |
| Tied to one AI provider | Works with any IDE agent |
| No version control | Git-native — diff, branch, PR your workflows |
| Hard to share or reuse | Clone, fork, compose |
An Agentfile workflow is a folder of plain text files your IDE agent reads and follows:
my-workflow/
workflow.yaml # Steps, goals, and routing logic
agents/ # Agent personas and rules (.md files)
skills/ # Reusable instruction sets (.md files)
scripts/ # Runtime scripts for CLI execution (optional)
outputs/ # Artifacts produced at each step (gitignored)
Your IDE agent loads these files and executes each step — no SDK, no framework, no lock-in. For CLI execution, scripts/run.sh and scripts/run.ps1 provide the orchestration layer.
npm install -g @tisonek/agentfile# Save your API key once
agentfile config set api-key your-anthropic-api-key
# Set default shell (optional - auto-detected by OS)
agentfile config set shell pwsh # or bash
# View current configuration
agentfile configmkdir my-project && cd my-project
agentfile init
agentfile create code-reviewerSlash Commands (Recommended)
/agentfile:run code-reviewer src/auth.js
/agentfile:create my-workflow Analyze logs and surface error patterns
/agentfile:list
Cursor / Windsurf — type any slash command directly in the composer.
Claude Code:
claude "/agentfile:run code-reviewer src/auth.js"Cline / Roo / GitHub Copilot — reference workflow.yaml in your system prompt or workspace instructions.
agentfile run code-reviewer --input "path/to/code.js"
# API key automatically loaded from configname: code-reviewer
version: 1.0.0
description: Reviews code for bugs, style, and improvements.
trigger:
type: natural-language
input_var: AGENT_INPUT
steps:
- id: analyze
name: Analyze Code
agent: agents/analyzer.md
skill: skills/code-analysis.md
input: $AGENT_INPUT
goal: Identify bugs, security issues, and code smells.
produces: outputs/01-analysis.md
- id: review
name: Write Review
agent: agents/reviewer.md
skill: skills/write-review.md
input: outputs/01-analysis.md
goal: Turn the analysis into a clear, actionable review report.
produces: outputs/02-review.md
gate: human-approval→ Full spec: SPEC.md → JSON Schema: schema/workflow.schema.json
| Command | Description |
|---|---|
agentfile init |
Scaffold a new Agentfile project in the current directory |
agentfile create <name> |
Create a new workflow |
agentfile run <name> |
Run a workflow by name |
agentfile list |
List all workflows in the current project |
agentfile validate [name] |
Validate workflow(s) against the spec |
agentfile config |
Manage configuration (API keys, default model, shell) |
agentfile setup-ide <ide> |
Generate IDE integration instructions for slash commands |
# Show current configuration
agentfile config
# Set API key (saved for future use)
agentfile config set api-key your-anthropic-api-key
# Set default model
agentfile config set model claude-sonnet-4-6
# Set default shell (overrides OS auto-detection)
agentfile config set shell pwsh # or bash
# Remove configuration
agentfile config unset api-key# Generate IDE-specific instructions for slash commands
agentfile setup-ide cursor # For Cursor
agentfile setup-ide windsurf # For Windsurf
agentfile setup-ide copilot # For GitHub Copilot
agentfile setup-ide claude # For Claude Code
agentfile setup-ide cline # For Cline
agentfile setup-ide roo # For Roo| Workflow | Description |
|---|---|
examples/hello-world |
Start here — the simplest possible workflow. One agent, one skill, no config needed |
examples/pr-summarizer |
Summarizes pull request diffs into structured reports |
workflows/code-reviewer |
Reviews code for bugs, security issues, and improvements |
workflows/slash-demo |
Slash command demo — perfect for testing /agentfile:run slash-demo hello |
workflows/workflow-creator |
Meta-workflow — generates new workflows from a natural language description |
Clone and point your IDE agent at examples/hello-world to get started in under a minute:
git clone https://github.com/TisoneK/agentfileThen in your IDE agent:
/agentfile:run hello-world recursion
Agentfile uses a single /agentfile: namespace with three subcommands:
/agentfile:run <workflow-name> <args>
/agentfile:create <new-workflow-name> <description>
/agentfile:list
Examples:
/agentfile:run hello-world recursion
/agentfile:run code-reviewer src/components/Button.js
/agentfile:run pr-summarizer https://github.com/user/repo/pull/123
/agentfile:create security-scanner Scan for OWASP vulnerabilities and produce a risk report
/agentfile:list
When an IDE agent sees an /agentfile: command, it:
- Parses the subcommand (
run/create/list) list— scansworkflows/*/workflow.yamland returns names + descriptions, no LLM neededcreate— invokesworkflow-creatorin IDE mode with the name and description pre-setrun— locatesworkflows/<name>/workflow.yaml, checksexecution.preferred, loads agents and skills, executes steps sequentially
→ Full IDE processing guide: docs/ide-slash-commands.md
| IDE | How to Use |
|---|---|
| Cursor | Type /agentfile:run name args in composer |
| Windsurf | Use slash command in cascade or rules |
| Claude Code | claude "/agentfile:run name args" |
| GitHub Copilot | Add to workspace instructions |
| Cline/Roo | Include in system prompt |
agentfile/
AGENTS.md # ← IDE agents read this first — slash command protocol
CLAUDE.md # Claude Code pointer → AGENTS.md
.cursorrules # Cursor pointer → AGENTS.md
.windsurfrules # Windsurf pointer → AGENTS.md
.clinerules # Cline/Roo pointer → AGENTS.md
SPEC.md # Formal specification
schema/
workflow.schema.json # JSON Schema (IDE autocomplete + validation)
examples/
hello-world/ # Example: minimal workflow
pr-summarizer/ # Example: PR summary workflow
workflows/
code-reviewer/ # Example: code review workflow
slash-demo/ # Demo: slash command testing
workflow-creator/ # Meta-workflow: generates new workflows
test-workflow/ # Test workflow for validation
ide-only/ # IDE-only execution demo
shared/
AGENTS.md # Global agent rules (injected in CLI calls)
project.md # Project-level conventions (injected in CLI calls)
cli/ # Node.js CLI source
docs/
concepts.md # Deep dive: how everything works
ide-slash-commands.md # Full slash command reference
~/.agentfile/
config.json # User configuration (API keys, defaults)
| Agent | How to use |
|---|---|
| Cursor | Paste workflow instructions into composer or .cursor/rules |
| Windsurf | Use cascade or add to rules file |
| GitHub Copilot | Add to workspace instructions |
| Claude Code | Pass workflow.yaml as task context |
| Cline | Reference in system prompt |
| Roo | Reference in system prompt |
- Windows: Auto-detects PowerShell with execution policy bypass
- macOS/Linux: Uses bash by default
- CLI Runtime: Requires
scripts/run.shandscripts/run.ps1 - IDE Agents: Execute steps directly (no scripts needed)
Configure default shell with agentfile config set shell <bash|pwsh>
Agentfile workflows support two execution modes with dual script support:
IDE Mode (Default)
- Follows
workflow.yamlsteps directly - Uses
scripts/ide/instructions for guidance - No external dependencies required
- Best for interactive development
CLI Mode
- Executes
scripts/cli/run.shorscripts/cli/run.ps1 - Scripts handle API calls and orchestration
- Required for automation/CI/CD
- Best for production workflows
Dual Script System
# workflow.yaml
execution:
preferred: "ide" # or "cli"Generated Structure:
scripts/
ide/ # IDE agent instructions
instructions.md # Step-by-step guide
steps.md # IDE-specific steps
cli/ # CLI runtime scripts
run.sh # Unix/Linux script
run.ps1 # Windows PowerShell
README.md # Execution documentation
IDE agents automatically detect the preferred mode and execute accordingly.
New examples, IDE adapter guides, and spec improvements are all welcome. See CONTRIBUTING.md.
git checkout -b my-feature
# make your changes
git commit -m "add: my-workflow example"
# open a pull requestMIT — see LICENSE