plutus-ai 0.3.203

Autonomous AI agent with subprocess orchestration, dynamic tool creation, and a local-first web interface

Plutus — Autonomous AI Agent with Subprocess Orchestration

A better, easier-to-use AI agent that spawns subprocesses to edit code, analyze files, and create new tools on the fly.

Features • Quick Start • Architecture • Tools • Dynamic Tools • Configuration

---

What is Plutus?

Plutus is an autonomous AI agent system that gives Claude (or any LLM) the ability to spawn isolated subprocesses for file editing, code analysis, shell execution, and dynamic tool creation. Think of it as Claude Code on steroids — the AI can not only run commands and edit files, but also create entirely new tools at runtime to solve problems it wasn't originally designed for.

Key Differentiators

Feature	OpenClaw	Plutus
File editing	Basic read/write	Subprocess-isolated surgical edits with diff output
Code analysis	None	Full AST analysis (functions, classes, complexity, call graphs)
Subprocess spawning	None	Parallel worker pool with JSON protocol
Dynamic tool creation	None	Create, validate, and hot-load new Python tools at runtime
CLI experience	Basic	Rich interactive REPL with slash commands
Guardrails	Basic	4-tier system (observer → autonomous) with audit logging
Planning	None	Built-in plan/step tracking with auto-progress

Features

Subprocess Orchestration

The agent spawns isolated worker subprocesses for every operation — file edits, code analysis, shell commands, and custom scripts all run in their own process with resource limits and timeouts.

Intelligent Code Editing

Surgical find/replace edits with diff output. The agent reads files, applies precise changes, and verifies the result — all in subprocess isolation.

Deep Code Analysis

AST-based analysis of Python files:

• Function and class extraction with signatures
• Cyclomatic complexity scoring (A–F ratings)
• Import dependency mapping
• Call graph generation
• TODO/FIXME/HACK detection
• Module summarization

Dynamic Tool Creation

The agent can write new Python tools at runtime:

1. Writes the tool code
2. Validates it (syntax check)
3. Saves it to ~/.plutus/custom_tools/
4. Hot-loads it into the tool registry
5. Uses it immediately

4-Tier Guardrail System

• Observer — Read-only, AI can only observe
• Assistant — Every action requires user approval
• Operator — Pre-approved actions run autonomously
• Autonomous — Full control, no restrictions

Multiple Interfaces

• Terminal REPL (plutus chat) — Rich interactive chat with slash commands
• Single prompt (plutus run "...") — Execute one task and exit
• Web UI (plutus start) — Full web interface with WebSocket streaming

Quick Start

Installation

# Clone the repository
git clone https://github.com/Crypt0nly/plutus.git
cd plutus

# Install with pip
pip install -e .

# Run setup wizard
plutus setup

First Run

# Interactive terminal chat
plutus chat

# Or run a single prompt
plutus run "Create a Python script that sorts a CSV file by the second column"

# Or launch the web UI
plutus start

Chat Commands

Inside plutus chat, use slash commands:

Command	Description
/help	Show available commands
/tools	List all available tools
/plan	Show current execution plan
/clear	Start a new conversation
/tier	Show or change guardrail tier
/workers	Show active subprocesses
/exit	Exit the chat

Architecture

┌─────────────────────────────────────────────────────┐
│                   Agent Runtime                      │
│  ┌──────────┐  ┌──────────┐  ┌──────────────────┐  │
│  │   LLM    │  │ Planner  │  │   Guardrails     │  │
│  │ (Claude) │  │          │  │ (4-tier system)  │  │
│  └────┬─────┘  └──────────┘  └──────────────────┘  │
│       │                                              │
│  ┌────▼──────────────────────────────────────────┐  │
│  │              Tool Registry                     │  │
│  │  ┌────────┐ ┌────────────┐ ┌───────────────┐  │  │
│  │  │ Shell  │ │ Code Editor│ │ Code Analysis │  │  │
│  │  └────────┘ └────────────┘ └───────────────┘  │  │
│  │  ┌────────────┐ ┌──────────────┐ ┌─────────┐  │  │
│  │  │ Subprocess │ │ Tool Creator │ │ Browser │  │  │
│  │  └────────────┘ └──────────────┘ └─────────┘  │  │
│  │  ┌──────────┐ ┌─────────┐ ┌─────────────────┐ │  │
│  │  │Filesystem│ │ Process │ │ Custom Tools... │ │  │
│  │  └──────────┘ └─────────┘ └─────────────────┘ │  │
│  └───────────────────┬───────────────────────────┘  │
│                      │                               │
│  ┌───────────────────▼───────────────────────────┐  │
│  │           Subprocess Manager                   │  │
│  │  ┌─────────────┐  ┌──────────────────────┐    │  │
│  │  │ Worker Pool │  │  JSON stdin/stdout    │    │  │
│  │  │ (max: 8)    │  │  protocol             │    │  │
│  │  └─────────────┘  └──────────────────────┘    │  │
│  └───────────────────────────────────────────────┘  │
│                      │                               │
│  ┌───────────────────▼───────────────────────────┐  │
│  │           Worker Subprocesses                  │  │
│  │  ┌──────┐ ┌──────────┐ ┌──────────┐ ┌──────┐ │  │
│  │  │Shell │ │File Edit │ │Code Anal.│ │Custom│ │  │
│  │  │Worker│ │Worker    │ │Worker    │ │Worker│ │  │
│  │  └──────┘ └──────────┘ └──────────┘ └──────┘ │  │
│  └───────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────┘

Subprocess Communication Protocol

Workers communicate via JSON over stdin/stdout (one JSON object per line):

Agent → Worker:  {"action": "edit", "path": "/file.py", "edits": [...]}
Worker → Agent:  {"success": true, "result": {"changes": 2, "diff": "..."}}

This design provides:

• Isolation — each operation runs in its own process
• Safety — crashes in workers don't affect the agent
• Parallelism — multiple workers can run simultaneously
• Simplicity — JSON protocol is easy to debug and extend

Tools

Built-in Tools

Tool	Description
shell	Execute shell commands
filesystem	File system operations (legacy, still available)
code_editor	Create, read, and edit files via subprocess
code_analysis	AST-based Python code analysis via subprocess
subprocess	Direct subprocess spawning for parallel tasks
tool_creator	Create new tools at runtime
process	System process management
system_info	System information queries
browser	Web browsing (Playwright)
clipboard	Clipboard operations
desktop	Desktop/window management
app_manager	Application management

Code Editor Operations

read       — Read file content (with optional line range)
write      — Create or overwrite a file
append     — Append content to a file
edit       — Apply surgical find/replace edits
delete     — Delete a file or directory
move       — Move/rename a file
copy       — Copy a file or directory
mkdir      — Create directories
list       — List directory contents
find       — Find files by glob pattern
grep       — Search file contents with regex
diff       — Show diff between two files

Code Analysis Operations

analyze        — Full analysis (everything below combined)
find_functions — List all function/method definitions with signatures
find_classes   — List all class definitions with methods
find_imports   — Extract all import statements
find_todos     — Find TODO/FIXME/HACK/NOTE comments
complexity     — Calculate cyclomatic complexity per function
symbols        — Extract all top-level symbols
call_graph     — Build function call graph
summarize      — Generate human-readable summary

Dynamic Tool Creation

The agent can create new tools when it encounters a task that requires capabilities it doesn't have:

# Example: Agent creates a CSV processor tool
tool_creator(
    operation="create",
    tool_name="csv_processor",
    description="Process and transform CSV files",
    code="""
import csv
from pathlib import Path

def main(args):
    path = args.get('path', '')
    operation = args.get('operation', 'read')
    
    if operation == 'read':
        with open(path) as f:
            reader = csv.DictReader(f)
            rows = list(reader)
        return {'success': True, 'result': {'rows': rows, 'count': len(rows)}}
    
    elif operation == 'sort':
        column = args.get('column', '')
        with open(path) as f:
            reader = csv.DictReader(f)
            rows = sorted(list(reader), key=lambda r: r.get(column, ''))
        return {'success': True, 'result': {'rows': rows, 'count': len(rows)}}
    
    return {'success': False, 'error': f'Unknown operation: {operation}'}
"""
)

Created tools are:

• Validated — syntax-checked before saving
• Persisted — saved to ~/.plutus/custom_tools/ across sessions
• Hot-loaded — immediately available in the tool registry
• Isolated — executed in subprocess workers

Configuration

Config File: ~/.plutus/config.json

{
  "model": {
    "provider": "anthropic",
    "model": "claude-sonnet-4-6",
    "temperature": 0.7,
    "max_tokens": 4096
  },
  "guardrails": {
    "tier": "operator",
    "audit_enabled": true
  },
  "agent": {
    "max_tool_rounds": 25
  },
  "planner": {
    "enabled": true,
    "auto_plan": true
  },
  "gateway": {
    "host": "127.0.0.1",
    "port": 7777
  }
}

Supported Providers

Provider	Models	Config
Anthropic	Claude 4 Sonnet, Claude 4 Opus, etc.	ANTHROPIC_API_KEY
OpenAI	GPT-4.1, GPT-4.1-mini, etc.	OPENAI_API_KEY
Ollama	Llama 3.2, Mistral, etc.	Local, no key needed
Custom	Any OpenAI-compatible endpoint	API_KEY + base URL

API Keys

Keys are stored securely in ~/.plutus/.secrets.json (chmod 600) and never exposed via the API. Set them via:

# Setup wizard
plutus setup

# Environment variable
export ANTHROPIC_API_KEY=sk-ant-...

# Or via the web UI settings page

CLI Reference

plutus                  # Show help
plutus start            # Launch web UI + API server
plutus chat             # Interactive terminal chat
plutus run "prompt"     # Run a single prompt
plutus setup            # Setup wizard
plutus status           # Show configuration
plutus tools            # List available tools
plutus set-tier <tier>  # Change guardrail tier
plutus audit            # Show audit log
plutus config-show      # Display full config as JSON

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run specific test file
pytest tests/test_subprocess.py -v

# Lint
ruff check plutus/

Project Structure

plutus/
├── plutus/
│   ├── __init__.py
│   ├── __main__.py
│   ├── cli.py                      # CLI with chat REPL
│   ├── config.py                   # Configuration management
│   ├── core/
│   │   ├── agent.py                # Main agent runtime
│   │   ├── conversation.py         # Conversation management
│   │   ├── heartbeat.py            # Heartbeat system
│   │   ├── llm.py                  # LLM client (LiteLLM)
│   │   ├── memory.py               # SQLite memory store
│   │   ├── planner.py              # Plan management
│   │   └── subprocess_manager.py   # Subprocess orchestrator
│   ├── gateway/                    # Web API + WebSocket
│   ├── guardrails/                 # Permission tiers + audit
│   ├── skills/                     # YAML skill definitions
│   ├── tools/
│   │   ├── base.py                 # Tool base class
│   │   ├── registry.py             # Tool registry with hot-reload
│   │   ├── code_analysis.py        # AST-based code analysis
│   │   ├── code_editor.py          # File creation and editing
│   │   ├── subprocess_tool.py      # Direct subprocess spawning
│   │   ├── tool_creator.py         # Dynamic tool creation
│   │   ├── shell.py                # Shell commands
│   │   ├── filesystem.py           # File system operations
│   │   ├── process.py              # Process management
│   │   ├── browser.py              # Web browsing
│   │   └── ...
│   └── workers/
│       ├── shell_worker.py         # Shell command worker
│       ├── file_edit_worker.py     # File editing worker
│       ├── code_analysis_worker.py # Code analysis worker
│       └── custom_worker.py        # Dynamic tool worker
├── ui/                             # React web interface
├── bridge/                         # Local ↔ cloud workspace sync
├── shared/                         # Shared models and memory (used by local + cloud)
├── tests/
│   ├── test_subprocess.py          # 34 comprehensive tests
│   ├── test_config.py
│   ├── test_guardrails.py
│   └── test_tools.py
├── pyproject.toml
└── README.md

Cloud Version

A hosted, fully managed version of Plutus is available at useplutus.ai for users who want a zero-setup cloud experience. The cloud version includes:

• Persistent cloud workspace with push/pull sync to your local Plutus
• Multi-device access via any browser
• Managed infrastructure — no API keys or server setup required
• Priority support

The cloud backend is closed-source and maintained in a private repository. If you are interested in self-hosting the full cloud stack, please get in touch.

License

MIT License — see LICENSE for details.