⚡ ToolNet CLI

The Autonomous Multi-Agent Terminal AI Coding Assistant

ToolNet CLI is a next-generation, high-performance terminal coding assistant engineered for developers, devops, and agentic workflows. Built on the Unified AgentHarness 2.0 architecture, it unifies interactive terminal streaming, autonomous multi-agent task graphs (DAG), role-based sub-agent delegation, intelligent context compaction, and multi-tier security sandboxing into a single developer tool.

Quick Start • Architecture • Features • Slash Commands • Keybindings • Security 2.0

🌟 Core Features

• ⚙️ Unified AgentHarness 2.0 Kernel: Central execution engine coordinating Context, Security, Tool Registry, MCP Gateway, and Session Persistence across all runtime modes.
• 🤖 Real Sub-Agent Execution Engine: Autonomous child agent loop with specialized personas (RESEARCHER, CODER, TESTER, REVIEWER, ARCHITECT, GENERAL), role-based tool filtering, and dependency chaining.
• 🧠 Unified Context Management Engine: Exact token estimation across model families, automatic bulky tool output pruning, atomic turn compaction (preserving assistant.tool_calls and role: "tool" pairs), and persistent session memory.
• 🛡️ Security & Permissions 2.0 (SecretGuard): 5-tier semantic command risk classification (CRITICAL_DENY, DANGEROUS, SAFE_BUILD, SAFE_READ, MODERATE_WRITE), sensitive file shielding (.env*, .ssh, .aws, .npmrc), and smart session trust ([A] Allow for Session).
• 🖥️ Native ANSI Full-Screen TUI: Zero-dependency TUI compatible with Termius, mobile SSH, tmux, and all standard terminal emulators. Supports bracketed paste, streaming markdown rendering, and live tool status indicators.
• 🔌 Model Context Protocol (MCP): Seamless integration with local and remote MCP servers via stdio and SSE.
• ⚡ Multiple Execution Strategies:
  • Interactive TUI: Live conversational pairing with streaming responses.
  • Headless Prompt (-p): Single-pass execution for scripts, CI/CD, and pipelines with optional --json output.
  • Turbo Mode: Ultra low-latency single-pass resolution for localized micro-tasks.
  • Teamwork DAG Orchestrator: Dynamic task graph decomposition executing multi-agent DAGs concurrently.

🚀 Quick Start

Installation

macOS / Linux (recommended)

curl -fsSL https://raw.githubusercontent.com/LBT-AI/Toolnet-CLI/main/install.sh | sh

This downloads a standalone binary — no Bun or Node.js required.

npm

npm install -g toolnetcli@latest

Windows (PowerShell)

irm https://raw.githubusercontent.com/LBT-AI/Toolnet-CLI/main/install.ps1 | iex

Homebrew / Scoop

Homebrew and Scoop manifests are maintained in this repository, but the public tap/bucket has not been published yet. Until registry publication is complete, use the standalone installer or npm methods above.

Uninstall

# Binary install
rm /usr/local/bin/toolnet  # or ~/.local/bin/toolnet

# npm install
npm uninstall -g toolnetcli

Launching Interactive Mode

# Launch interactive TUI in the current workspace
toolnet

# Resume the most recent session
toolnet --resume

# Load a specific session ID
toolnet --session sess_12345

Non-Interactive (Headless) Mode

# Run a single query directly from CLI / scripts
toolnet -p "Audit this repository for security vulnerabilities and summarize dependencies"

# Structured JSON output for CI/CD automation
toolnet -p "Check for TypeScript errors in src/" --json

🏗️ Architecture

┌────────────────────────────────────────────────────────────────────────────┐
│                        UNIFIED AGENT HARNESS 2.0                           │
│                 (Master Execution Kernel & Orchestration)                  │
├────────────────────────────────────────────────────────────────────────────┤
│                                                                            │
│  [Runtime Context]        [Core Subsystems]        [Execution Modes]       │
│  ├── Workspace & Root     ├── Context Engine       ├── Interactive (TUI)   │
│  ├── Session Persistence  ├── Security 2.0 Guard   ├── Headless (-p)       │
│  ├── Model Gateway        ├── Tool Registry & MCP  ├── Turbo Single-Pass   │
│  └── Framework Detector   └── EventBus Telemetry   └── Teamwork DAG        │
│                                                                            │
└─────────────────────────────────────┬──────────────────────────────────────┘
                                      │
           ┌──────────────────────────┴──────────────────────────┐
           ▼                                                     ▼
┌──────────────────────────────┐              ┌──────────────────────────────┐
│       MAIN AGENT RUNTIME     │              │    SUB-AGENT RUNTIME ENGINE  │
│  • Autonomous Tool Loop      │─── delegates ├── Researcher (Read/Grep/Web) │
│  • Interactive Streaming     │     via      ├── Coder (Write/Patch/Edit)   │
│  • Multi-turn Session Flow   │ spawn_agent  ├── Tester (Test/Diagnostics)  │
│  • Atomic Compactor          │              ├── Reviewer (Diff/Security)   │
└──────────────────────────────┘              └──────────────────────────────┘

⌨️ Interactive Keybindings

| Keybinding | Action | Description | | :--- | :--- | :--- | | Enter | Send Message | Submits your prompt to the active agent | | Tab | Toggle Mode | Switches between Build mode (code execution) and Plan mode (read-only planning) | | Ctrl+K | Switch Model | Opens the interactive model & provider picker | | Ctrl+C | Cancel / Exit | Aborts the active generation or exits the TUI | | Ctrl+L | Clear Screen | Clears the terminal screen buffer | | PageUp / PageDown | Scroll Chat | Navigates through long conversation history |

💬 Slash Commands

| Command | Aliases | Description | | :--- | :--- | :--- | | /harness | /kernel, /sys | Displays Unified AgentHarness status, active subsystems, framework detection, and token telemetry | | /subagent [role] <task> | /sub, /agent | Spawns a dedicated subagent (RESEARCHER, CODER, TESTER, REVIEWER, ARCHITECT) | | /sandbox [mode] | /sb | Inspects or sets sandbox mode (workspace, ask, full-access, or clear session rules) | | /compact | /compress | Triggers atomic context compaction and displays context token budget utilization | | /status | /st | Checks ToolNet API Gateway connectivity, active providers, and tunnels | | /model [name] | /m | Switches the active language model or lists available models | | /teamwork <prompt> | /tw | Decomposes a complex objective into an autonomous DAG task graph | | /tools | — | Lists all available tools, file operations, and execution permissions | | /mcp | — | Manages Model Context Protocol (MCP) servers and client tools | | /skills | — | Lists locally installed project skills and extensions | | /session [list\|new\|switch] | — | Manages multi-session persistence and history switching | | /undo / /redo | — | Reverts or reapplies file changes made by surgical patch tools | | /qa <prompt> | — | Runs autonomous automated QA and test suite diagnostics | | /doctor | — | Runs local diagnostic checks on environment, tools, and connectivity | | /update | — | Checks for ToolNet CLI updates on npm |

🛡️ Security & Permissions 2.0

ToolNet CLI includes an enterprise-grade, multi-layered security engine to keep your machine safe:

[Tool Call Request]
        │
        ▼
[SecretGuard] ───► Is sensitive file (.env, id_rsa, .npmrc, ~/.ssh)? ──► BLOCK in workspace mode
        │
        ▼
[Semantic Command Classifier] ───► Evaluate Shell AST (CRITICAL_DENY / DANGEROUS / SAFE)
        │
        ▼
[Session Trust Manager] ───► Was this tool/command allowed for the session? ──► ALLOW
        │
        ▼
[Interactive Approval Modal] ──► [Y] Once   [A] Allow for Session   [N] Deny
        │
        ▼
[Audit Logger] ───► Log structured JSONL event to .logs/security-audit.jsonl

Sandbox Modes

• workspace (Default): Restricts read/write operations strictly within the project workspace root. Blocks access to credential files (.env*, id_rsa*) and dangerous shell commands (rm -rf /, mkfs).
• ask: Prompts the user interactively before executing sensitive file operations or dangerous commands.
• full-access: Unrestricted execution mode for automated containerized CI/CD environments.

📂 Configuration & Storage

| Path | Purpose | | :--- | :--- | | ~/.toolnetcli/config.json | Canonical CLI configuration (gateway/API mode, default model, sandbox, theme, update checks) | | ~/.toolnetcli/sessions/ | Saved JSON session histories and message turns | | ~/.toolnetcli/config.json | Global CLI preferences (default model, base URL, provider, theme) | | .toolnet/permissions.json | Project-specific workspace security policy | | .logs/security-audit.jsonl | Structured security evaluation audit log |

Environment Variables

• TOOLNETCLI_SESSIONS_DIR: Custom path for session storage.
• TOOLNET_API_URL: Optional custom ToolNet Gateway endpoint URL.
• TOOLNET_DEBUG: Set to 1 to enable verbose debugging logs.

🛠️ Development & Testing

# Clone repository
git clone https://github.com/LBT-AI/Toolnet-CLI.git
cd Toolnet-CLI

# Install dependencies
bun install

# Full production verification: typecheck + 400+ tests + Bun/Node builds
bun run check

# Or run individual stages
bun run typecheck
bun test
bun run build

CI additionally smoke-tests the CLI and Teamwork persistence on Node 22 and the minimum supported Node 20 runtime.

Note — Node 20–21 persistence: native SQLite (node:sqlite) requires Node 22.5+. On Node 20–21, Teamwork checkpoints and context-cache stores automatically fall back to the built-in JSON persistence store (<db>.json next to the requested path). The same data model is preserved and nothing crashes — only raw SQL throughput differs.

📄 License

This project is licensed under the MIT License.