frago-cli 0.28.0

Frago - AI-driven multi-runtime automation framework with Chrome CDP, Python, and Shell support

frago - Multi-Runtime Automation Infrastructure

简体中文

Docs: Key Concepts · Installation · User Guide · Recipes · Architecture · Use Cases · Development

Quick Start

uv tool install frago-cli   # Install frago
frago init                   # Initialize environment
frago server start           # Start web service
# Open http://127.0.0.1:8093 in your browser

New to uv or setting up a fresh system? See the Installation Guide for prerequisites.

---

Manifesto

AI should free people from repetitive labor, not become a new instrument of extraction.

Three beliefs that guide frago:

1. Delivery over Dialogue

Chatting with AI produces nothing. The ICQ era of AI — endless conversation, zero delivery — wastes your time and money.

frago exists for results: recipes that run, scripts that execute, data that's extracted. If AI can't hand you a deliverable, it hasn't done its job.

2. Your Tools, Your Control

We reject the narrative that you must wait for some company to build AGI before automation serves you.

frago is open source. Your recipes, your skills, your Git repo. You accumulate capability, not subscription fees. The tools you build are yours — portable, version-controlled, independent.

3. Against Token Exploitation

Many "AI products" are token vending machines wrapped in pretty UIs. You pay per conversation, per generation, per retry — and get nothing persistent in return.

frago attacks this directly:

• First exploration: ~150k tokens
• Every run after: ~2k tokens (98.7% saved)

The savings compound. The recipes stay. Your time returns to family, hobbies, creation — not to feeding another revenue stream.

---

Recent Updates

Version	Highlights
v0.26.0	Workspace file browser; frago view media support (video, image, audio, 3D models)
v0.24.0	Community recipes infrastructure; recipe install/uninstall/update/search/share commands
v0.23.0	WebSocket real-time sync; YouTube recipes (download, subtitles, transcript)
v0.22.0	Cross-platform autostart; frago autostart command; integrated into init workflow
v0.21.0	i18n support; user language preference for AI title generation

Multi-runtime automation infrastructure designed for AI agents, providing persistent context management and reusable Recipe system.

---

Why frago

When facing prompts, AI can only "talk" but not "do"—it "talks once" but never "follows through from start to finish." Think of ChatGPT in 2023. So people designed Agents. Agents call tools through standardized interfaces.

But reality is: tasks are infinite, while tools are finite.

You ask AI to extract YouTube subtitles. It spends 5 minutes exploring, succeeds. The next day, same request—it starts from scratch again. It completely forgot what it did yesterday.

Even an Agent like Claude Code appears clumsy when facing each person's unique task requirements: every time it must explore, every time it burns through tokens, dragging the LLM from start to finish. Slow and unstable: out of 10 attempts, maybe 5 take the right path, while the other 5 are filled with "strange" and "painful" trial-and-error.

Agents lack context—that's a fact. But what kind of context do they lack?

People tried RAG, fragmenting information so Agents could retrieve and "find methods." This is "theoretically correct but practically misguided"—a massive pitfall. The key issue: each person's task requirements are "local" and bounded. They don't need a heavyweight RAG system. RAG over-complicates how individuals solve problems.

Research from Anthropic and Google both point to: directly consulting documentation. The author of this project proposed the same view in 2024. But this approach requires Agents with sufficient capability. Claude Code is exactly such an Agent.

Claude Code designed a documentation architecture: commands and skills, to practice this philosophy. frago builds on this foundation, deeply implementing the author's design philosophy: every piece of methodological knowledge must be tied to concrete executable tools.

In frago's framework, skills are collections of methodologies, and recipes are collections of executable tools.

The author's vision: through frago's Claude Code slash commands (/frago.run and other core commands), establish an Agent specification—enabling it to explore unfamiliar problems and standardize results into structured information; through self-awareness, proactively build the association between skills and recipes.

Ultimately, your Agent can fully understand your descriptions of work and task requirements, leverage existing skills to find and properly use relevant recipes, achieving "driving automated execution with minimal token cost."

frago is not the Agent itself, but the Agent's "skeleton."

Agents are smart enough, but not yet resourceful. frago teaches them to remember how to get things done.

---

How to Use

frago integrates with Claude Code through four slash commands, forming a complete "explore → solidify → execute" loop.

/frago.run     Explore and research, accumulate experience
     ↓
/frago.recipe  Solidify experience into reusable recipes
/frago.test    Validate recipes (while context is fresh)
     ↓
/frago.exec    Execute quickly with skill guidance

Step 1: Explore and Research

In Claude Code, type:

/frago.run Research how to extract YouTube video subtitles

The Agent will:

• Create a project to store this run instance
• Use frago's basic tools (navigate, click, exec-js, etc.) to explore
• Automatically record execution.jsonl and key findings
• Persist all screenshots, scripts, and output files

projects/youtube-transcript-research/
├── logs/execution.jsonl    # Structured execution logs
├── screenshots/            # Screenshot archive
├── scripts/                # Validated scripts
└── outputs/                # Output files

Step 2: Solidify Recipes

After exploration, type:

/frago.recipe

The Agent will:

• Analyze the experience accumulated during exploration
• Auto-generate necessary recipes for this task
• Create corresponding skills (coming soon)
• Associate skills with recipes

Generated recipe example:

---
name: youtube_extract_video_transcript
type: atomic
runtime: chrome-js
description: "Extract complete transcript text from YouTube videos"
use_cases:
  - "Batch extract video subtitle content for text analysis"
  - "Create indexes or summaries for videos"
---

Step 3: Validate Recipes

While the session context is still fresh, test immediately:

/frago.test youtube_extract_video_transcript

Validation failed? Fix it on the spot, no need to re-explore. This is why recipe and test should be parallel—debugging costs more after context is lost.

Step 4: Quick Execution

Next time you have a similar need, type:

/frago.exec video-production Create a short video about AI

The Agent will:

• Load the specified skill (video-production)
• Follow the methodology in the skill to invoke relevant recipes
• Complete the task quickly, no repeated exploration

This is the value of the "skeleton": 5 minutes to explore the first time, seconds to execute thereafter.

---

Technical Foundation

The above workflow relies on frago's underlying capabilities:

Capability	Description
Native CDP	Direct Chrome DevTools Protocol connection, ~2MB lightweight, no Node.js deps
Run System	Persistent task context, JSONL structured logs
Recipe System	Metadata-driven, three-tier priority (Project > User > Example)
Web Service	FastAPI backend + React frontend, browser-based GUI on port 8093
Multi-Runtime	Chrome JS, Python, Shell runtime support

Architecture Comparison:
Playwright:  Python → Node.js relay → CDP → Chrome  (~100MB)
frago:       Python → CDP → Chrome                  (~2MB)

---

frago Is Not Playwright/Selenium

Playwright and Selenium are testing tools—launch browser, run tests, close browser. Every run starts fresh.

frago is the skeleton for AI—connect to an existing browser, explore, learn, remember. Experience accumulates.

You need...	Choose
Quality assurance, regression testing, CI/CD	Playwright/Selenium
Data collection, workflow automation, AI-assisted tasks	frago
One-off scripts, run and discard	Playwright/Selenium
Accumulate experience, faster next time	frago

Technical differences (lightweight, direct CDP, no Node.js dependency) are outcomes, not goals.

The core difference is design philosophy: testing tools assume you know what to do; frago assumes you're exploring, and helps you remember what you discovered.

frago vs Dify/Coze/n8n

Dify, Coze, and n8n are workflow orchestration tools.

Traditional usage: manually drag nodes, connect lines, configure parameters. n8n launched AI Workflow Builder that can generate workflow nodes from natural language (Dify and Coze don't have similar features yet).

But whether manual or AI-assisted, what do you end up with? A flowchart.

Then what?

1. You still need to enter the platform, understand the diagram
2. Run, error, go back and modify node config
3. Run again, another error, modify again
4. After debugging passes, the flowchart runs

AI drew the diagram for you, but debugging, modifying, maintaining—still your job.

Using frago:

/frago.run Scrape data from this website

No flowchart. AI goes to work directly—opens browser, clicks, extracts data, handles errors. You just wait.

When done:

/frago.recipe

Recipe auto-generated. Next time:

/frago.exec Scrape similar website

You don't need to enter any platform, don't need to look at any flowchart.

	Orchestration Tools (incl. AI-assisted)	frago
What AI does	Draws flowcharts for you	Does the work directly
What you do	Enter platform, read diagrams, debug, modify config	State needs, wait for results
Output	A flowchart that needs maintenance	Reusable recipe

Orchestration tools' AI is your "diagram assistant"; frago's AI is your "executor".

Of course, if you need scheduled triggers, visual monitoring, team collaboration approvals—orchestration tools are better fits. But if you just want to get things done—frago lets you solve problems by talking, no platform to learn.

---

Resource Management

Why Resource Sync Commands

frago is open-source—anyone can install it via PyPI. But the skeleton is universal, while the brain is personal.

Each person has:

• Their own application scenarios
• Personalized knowledge (skills)
• Custom automation scripts (recipes)

These personalized resources shouldn't live in the public package. They belong to you.

frago's philosophy: cross-environment consistency. Your resources should be available wherever you work—different machines, fresh installations, or new projects. The tool comes from PyPI; your brain comes from your private repository.

frago doesn't provide community-level cloud sync services (yet). Instead, it gives you commands to manage sync with your own Git repository.

Resource Flow Overview

┌─────────────┐   publish   ┌─────────────┐    sync    ┌─────────────┐
│   Project   │ ──────────→ │   System    │ ─────────→ │   Remote    │
│  .claude/   │             │ ~/.claude/  │            │  Git Repo   │
│  examples/  │             │ ~/.frago/   │            │             │
└─────────────┘             └─────────────┘            └─────────────┘
       ↑                          │                          │
       │       dev-load           │         deploy           │
       └──────────────────────────┴──────────────────────────┘

Commands

Command	Direction	Purpose
publish	Project → System	Push project resources to system directories
sync	System → Remote	Push system resources to your private Git repo
deploy	Remote → System	Pull from your private repo to system directories
dev-load	System → Project	Load system resources into current project (dev only)

Typical Workflows

Developer Flow (local changes → cloud):

# After editing recipes in your project
frago publish              # Project → System
frago sync                 # System → Remote Git

New Machine Flow (cloud → local):

# First time setup on a new machine
frago sync --set-repo [e-mail]:you/my-frago-resources.git
frago deploy               # Remote Git → System
frago dev-load             # System → Project (if developing frago)

Regular User (just uses frago):

frago deploy               # Get latest resources from your repo
# Resources are now in ~/.claude/ and ~/.frago/, ready to use

What Gets Synced

Only frago-specific resources are synced:

• frago.*.md commands (not your other Claude commands)
• frago-* skills (not your other skills)
• All recipes in ~/.frago/recipes/

Your personal, non-frago Claude commands and skills are never touched.

---

Documentation Navigation

• Key Concepts - Skill, Recipe, Run definitions and relationships
• Use Cases - Complete workflow from Recipe creation to Workflow orchestration
• Architecture - Core differences, technology choices, system design
• Installation - Installation methods, dependencies, optional features
• User Guide - CDP commands, Recipe management, Run system
• Recipe System - AI-First design, metadata-driven, Workflow orchestration
• Development - Project structure, development standards, testing methods
• Roadmap - Completed features, todos, version planning

Writings

Personal thoughts on AI automation, Agent design, and lessons learned.

→ Read the Writings

---

Project Status

📍 Current Stage: Full-featured workspace with media preview support

Latest Features (v0.17.0 - v0.26.0):

• ✅ Workspace file browser - Browse run instance directories in Web UI
• ✅ Media viewer - frago view supports video, image, audio, 3D models (glTF/GLB)
• ✅ Community recipes - recipe install/uninstall/update/search/share for community contributions
• ✅ WebSocket real-time sync - Server push updates, reduced polling
• ✅ YouTube recipes - Download videos, extract subtitles and transcripts
• ✅ Cross-platform autostart - frago autostart manages server boot startup (macOS/Linux/Windows)
• ✅ i18n support - UI internationalization with user language preferences
• ✅ Web service mode - frago server launches browser-based GUI on port 8093

Core Infrastructure:

• ✅ Native CDP protocol layer (direct Chrome control, ~2MB lightweight)
• ✅ Recipe metadata-driven architecture (chrome-js/python/shell runtime)
• ✅ Run command system (topic-based task management, JSONL structured logs)
• ✅ Web service backend (FastAPI + React frontend)
• ✅ CLI tools and grouped command system

See Roadmap for details

---

License

AGPL-3.0 License - see LICENSE file

Contributing

Issues and Pull Requests are welcome!

• Project issues: Submit Issue
• Technical discussion: Discussions

Contributors

---

Created with Claude Code | 2025-11