ScopeForge

Commercial-Grade Autonomous AI Agent Harness & Coding Assistant
Claude Code & Antigravity Style Terminal Assistant, LangGraph Multi-Agent Team, LangChain Tool Harness, Real-Time Hardware Resource Telemetry, Persistent Session Management & Specialized Cybersecurity Swarms.

⚡ What is ScopeForge?

ScopeForge is a commercial-grade autonomous AI coding assistant and agent harness designed to compete directly with Claude Code, Antigravity, and Open Code. It operates seamlessly as a full-featured general developer harness for software engineering, code refactoring, system automation, web research, and tool orchestration—while offering specialized built-in cybersecurity swarms (ScopeGate RoE sandbox, SAST code auditing, port scanning, and falsifiable proof-of-concept verification) as an advanced capability.

🌟 Key Commercial Agent Features:

• Claude Code & AGY Native Scrollback Interface: Runs cleanly in your terminal without screen clobbering, with real-time markdown streaming, dynamic tool cards, and two-line prompt outline.
• Hardware & Device Awareness: Real-time telemetry displaying CPU core count, system RAM capacity, platform architecture, and running token / dollar cost estimations.
• Session Persistence & Context Management (/resume, /context, /history): Automatically persists sessions to ~/.scopeforge/sessions/, allows listing and resuming past conversations seamlessly, and provides full context inspection.
• Flexible Memory & Extensibility Control (/clear, /reset, /reload): Reset memory, conversation context, rescan workspace skills, or reconnect MCP bridges on-the-fly.
• Unpacked & Compact Tool Output Cards: Automatically unwraps JSON outputs from bash commands and tools, caps verbose logs to 6 lines max with clean expansion, and eliminates terminal clutter.

🚀 Installation & Quickstart

1. Install via pip / pipx / uv

# Recommended: Install using pipx (isolated environment)
pipx install scopeforge

# Or standard pip
pip install scopeforge

# Or run instantly without installation via uv
uvx scopeforge

2. Launch the Interactive TUI

Simply type scopeforge or the shorthand sf in your terminal:

scopeforge
# Or:
sf

ScopeForge will launch directly into the dark-themed reactive terminal UI.

⌨️ Modern Terminal Shortcuts

ScopeForge uses standard modern terminal Ctrl+[key] bindings:

🤖 Multi-Agent Team (LangGraph)

ScopeForge routes tasks dynamically to specialized autonomous agents:

• 🧠 Supervisor: General architecture, system management, code discussion, and high-level strategy.
• 💻 DevAgent: Local code workspace operations, terminal sandbox commands, skill installation, and live web research.
• 🛰️ ReconAgent: Passive & active reconnaissance, port discovery telemetry, and attack surface enumeration.
• 🔬 AuditAgent: Static code auditing (SAST), CVE advisory correlation, and OWASP vulnerability identification.
• ⚡ ExploitAgent: Controlled proof-of-concept (PoC) validation within strict authorized bounds.
• 📋 ReportAgent: Executive SecOps summary, CVSS scoring, and remediation action plans.

Agents communicate asynchronously using the cryptographically verified Agent-to-Agent (A2A) Protocol Bus.

🛠️ LangChain Autonomous Tool-Calling Harness

ScopeForge provides real autonomous tool execution through LangChain:

• Autonomous Tool Binding: Models (OpenRouter, Groq, DeepSeek, Claude, OpenAI) are automatically bound to the complete toolsuite and active MCP tools via chat_model.bind_tools().
• Iterative Tool Execution: The harness detects tool_calls, executes tools safely via the ScopeGate middleware pipeline, feeds ToolMessage outputs back into context, and synthesizes answers iteratively.
• Built-in Tools:
  • bash_cli: Sandboxed shell execution with dangerous-pattern blocking.
  • google_web_search: Live web search for fast technical lookups.
  • view_file / write_file / edit_file: Surgical code workspace inspection and editing.
  • glob_files / grep_search: Fast repository pattern discovery.
  • git_status_tool / git_diff_tool / git_commit_tool: Git version control operations.
  • recon_port_scan / web_surface_probe: Perimeter discovery tools.
  • sast_code_audit / cve_advisory_search: Vulnerability triage tools.

🌐 Model Support & Custom API Configuration

ScopeForge supports any OpenAI-compatible, Anthropic, or Ollama provider:

Quick Presets (Zero Setup):

• openrouter-free: Meta-router for high-availability free frontier models (openrouter/free).
• openrouter-free-nemotron: NVIDIA Nemotron 550B Free.
• groq-llama3: Llama 3.3 70B on Groq Ultra-Fast inference.
• ollama-llama3: Offline local models via Ollama.

Custom Endpoints & Third-Party Proxies:

Open the Model Switcher (Ctrl+M) or use slash commands:

/model
/config set key <API_KEY>
/config set model <MODEL_ID>
/config set base <CUSTOM_URL>

ScopeForge features native paste de-duplication and custom endpoint normalization (no forced path appending).

⇄ Agent-to-Agent (A2A) Protocol Bus & Subagents

ScopeForge implements first-class Agent-to-Agent collaboration inspired by modern autonomous multi-agent architectures:

• Dynamic Subagent Invocation: Use invoke_subagent(agent_name, task) to delegate tasks to recon, audit, exploit, report, or dev.
• A2A Message Broadcasting: Use send_a2a_message(recipient, intent, message) to broadcast evidence, handovers, and results across the bus.
• Interactive Telemetry:

  /a2a                                    # View registered agents, bus status, and recent message table
  /a2a send <agent> <intent> <message>    # Manually dispatch an A2A message
  /agent <name>                           # Direct next task to a specific specialist agent
  

🔌 Model Context Protocol (MCP) & Extensible Skills

Model Context Protocol (MCP):

ScopeForge connects directly to external MCP servers:

/mcp                                    # List all configured MCP servers & status
/mcp tools                              # View active tools registered across all MCP servers
/mcp add <name> <command>               # Add and immediately enable an MCP server
/mcp remove <name>                      # Remove an MCP server
/mcp enable <name>                      # Enable an existing server
/mcp disable <name>                     # Disable an existing server

Models can also invoke add_mcp_server and list_mcp_servers autonomously as LangChain tools.

Extensible Skills (SKILL.md):

Install custom domain playbooks and guidelines directly from GitHub:

/skill install https://github.com/Jakeschincariol/linkedin-agent-skill.git
/skill add my-playbook [description]    # Create a new custom skill template
/skill list                             # Discover all available local skills
/skill <name>                           # Toggle skill activation

Skills are automatically cloned, loaded into .scopeforge/skills/ and ~/.scopeforge/skills/, and dynamically injected into the model prompt when relevant keywords are triggered.

⚡ Long-Running Multi-Step Task Execution

Unlike basic wrappers that terminate prematurely after 2-4 tool calls, ScopeForge is built for real developer and security tasks:

• Up to 25 Autonomous Iterations: Run complete multi-step workflows (e.g. cloning a repository, installing skills, writing code, running tests, fixing errors, and auditing) without premature termination.
• Graceful Synthesis Safeguard: If an iteration budget is reached, ScopeForge automatically synthesizes all tool observations and outputs into a coherent, comprehensive final answer rather than cutting off in the middle.
• Clean Token Streaming: Token streaming seamlessly transitions between text generation and tool execution callouts with zero dropped chunks or duplicate messages.

🛡️ ScopeGate Rules of Engagement (RoE)

ScopeForge operates under 3 strict enforcement modes:

1. PLAN (plan): Zero network egress. Generates falsifiable hypotheses, models attack surface, and plans audits.
2. ARTIFACTS (artifacts): Inspects locally supplied source code, HAR files, and logs without external traffic.
3. LIVE (live): Network tools enabled strictly against authorized scope domains and IPs.

📜 License

Licensed under the Apache License, Version 2.0.