opensculpt 0.1.6

OpenSculpt — The Self-Evolving Agentic OS

The Self-Evolving Agentic OS

Every failure is a chisel strike. Every fix reveals a better shape. 🐧🪨

OpenSculpt is not another agent framework. It's an operating system for intelligence. You speak naturally, it reasons, plans, spawns agents, executes, remembers, and evolves — driven by what you actually need, not random research papers. Claude Code, Cursor, Windsurf are the chisels. The OS is the stone. Like Tux represents Linux, Chip (the self-sculpting penguin) represents an OS that shapes itself.

sculpt "why is my API slow?"
# → Spawns analyst agent, profiles endpoints, checks logs, reports findings

sculpt "write a REST API for user management with tests and docs"
# → Assembles a team: architect → coder → [tester + documenter in parallel]

Try It Free

No install needed — explore OpenSculpt directly in your browser:

Method	What you get	Time to start
	Full dev environment + dashboard auto-launches on port 8420	~60s
	Same setup via Gitpod	~90s

GitHub Codespaces gives 120 free core-hours/month on every GitHub account — enough to explore for days.

Once the Codespace starts, the dashboard opens automatically. Set your API key in the terminal to unlock the full OS:

export SCULPT_LLM_API_KEY=your-key-here
sculpt "analyze this codebase"

Install

pip

pip install opensculpt

From Source

git clone https://github.com/opensculpt/opensculpt.git
cd opensculpt
pip install -e ".[dev]"

Note: If sculpt isn't found after install, use python -m agos instead.

Windows Executable

Download OpenSculpt.exe from the latest release.

Quick Start

# Set your API key (Anthropic, OpenRouter, or any supported provider)
export SCULPT_LLM_API_KEY=your-key-here

# Initialize workspace
sculpt init

# Talk to it via CLI
sculpt "analyze my codebase and summarize what each module does"
sculpt "find all TODO comments and prioritize them"

# Or launch the web dashboard
python -m agos.serve
# Opens http://localhost:8420

Architecture

+------------------------------------------------------------------+
|                         OpenSculpt                                  |
|                                                                   |
|   INTERFACE    Natural Language CLI  |  Dashboard  |  SDK         |
|                         |                                         |
|   SOUL         Intent Engine (understand → plan → execute)        |
|                         |                                         |
|   BRAIN        Agent Kernel (lifecycle, state, budget)            |
|                         |                                         |
|   MEMORY       Knowledge System (episodic + semantic + graph)     |
|                         |                                         |
|   BODY         Tool Bus (file, shell, HTTP, Python, web search)   |
|                         |                                         |
|   SENSES       Triggers (file watch, cron, webhooks)              |
|                         |                                         |
|   SOCIAL       Coordination (channels, teams, debate protocol)    |
|                         |                                         |
|   IMMUNE       Policy Engine + Audit Trail                        |
|                         |                                         |
|   EVOLUTION    Demand-Driven Self-Evolution (failures → research  |
|                → code generation → sandbox → deploy)              |
+------------------------------------------------------------------+

CLI Commands

Command	Description
sculpt "<intent>"	Natural language — the OS figures out what to do
sculpt ps	List running agents
sculpt recall "<topic>"	Search knowledge system
sculpt timeline	View event history
sculpt watch <path> "<intent>"	Watch files, trigger agent on changes
sculpt schedule <interval> "<intent>"	Run agent on a schedule
sculpt team "<task>"	Multi-agent team execution
sculpt evolve	Run R&D cycle (scan arxiv, analyze, propose)
sculpt evolve --proposals	View pending evolution proposals
sculpt ambient --start	Start background watchers
sculpt proactive --scan	Run pattern detection
sculpt audit	View audit trail
sculpt policy	Configure safety policies
sculpt dashboard	Launch web monitoring UI
sculpt update	Check for updates and self-update
sculpt version	Show version

Configuration

All settings via environment variables with SCULPT_ prefix:

Variable	Default	Description
SCULPT_LLM_API_KEY	(required)	Your LLM API key (Anthropic, OpenRouter, etc.)
SCULPT_DEFAULT_MODEL	anthropic/claude-haiku-4-5	Claude model to use
SCULPT_WORKSPACE_DIR	.opensculpt	Local workspace directory
SCULPT_MAX_CONCURRENT_AGENTS	50	Max agents running at once
SCULPT_DASHBOARD_PORT	8420	Dashboard web UI port
SCULPT_LOG_LEVEL	INFO	Logging level

Self-Evolution

OpenSculpt evolves itself based on what you actually need:

1. Detect — Collects demand signals from user failures, tool errors, missing capabilities, agent crashes
2. Scout — Searches arxiv for papers that address actual problems (not random topics)
3. Analyze — Extracts actionable techniques from each paper
4. Generate — LLM writes code implementing the technique, with demand context injected
5. Test — Runs code patterns in a sandboxed environment
6. Deploy — Applies sandbox-passing code as live evolved strategies
7. Learn — Records what worked and what failed for cross-cycle learning

sculpt evolve                    # Run a full R&D cycle
sculpt evolve --proposals        # Review what it found
sculpt evolve --accept <id>      # Accept a proposal
sculpt evolve --apply <id>       # Apply it (with auto-rollback on failure)

Every failure is a chisel strike. Every fix reveals a better shape. The OS sculpts itself — shaped by AI coding tools on every user's machine.

Security

OpenSculpt is an operating system — it executes shell commands, writes files, makes HTTP requests, and manages Docker containers by design. This is equivalent to giving it sudo access.

Before exposing to a network:

• Set SCULPT_DASHBOARD_API_KEY to a strong random value
• Set SCULPT_APPROVAL_MODE to confirm-dangerous (default is auto)
• Do not expose port 8420 to the public internet without auth

By design, the OS agent can:

• Execute arbitrary shell commands and Python code
• Read/write any file on the filesystem
• Make HTTP requests to any URL
• Manage Docker containers via the Docker socket

This is intentional — it's an OS, not a sandboxed app. Treat it like SSH access to a server.

Development

pip install -e ".[dev]"
pytest tests/ -v          # 779+ tests
ruff check agos/ tests/   # lint

License

Apache License 2.0. See LICENSE.