Agent Engine

Ghost includes a full agent engine that can reason, plan, and execute multi-step tasks using local AI models.

ReAct Loop

The agent follows a Reason + Act cycle:

User request
    ↓
LLM reasons about the task
    ↓
LLM selects a tool to call
    ↓
Tool executes and returns results
    ↓
LLM reasons about results
    ↓
(repeat up to 10 iterations)
    ↓
Final answer to user

Built-in Tools

Tool	Description	Safety Tier
search_files	Hybrid search across indexed documents	Safe
read_file	Read file contents	Moderate
list_directory	List files in a directory	Safe
get_system_info	Hardware and platform info	Safe
get_index_stats	Database statistics	Safe
web_search	Web search via MCP server	Moderate

Plus unlimited external tools via connected MCP servers.

Safety System

Ghost uses a 3-tier risk classification:

Safe (Auto-approved)

• Read-only operations
• Search, list, stats queries
• No user confirmation needed

Moderate (Logged)

• File reads, especially sensitive paths
• External API calls via MCP
• Logged for transparency

Dangerous (Requires Approval)

• File writes, deletes, or moves
• Command execution
• System modifications
• Always asks for user confirmation

Conversation Memory

All conversations are persisted in SQLite:

• Full message history with timestamps
• FTS5 search across past conversations
• CASCADE deletes for clean conversation management
• Conversation summaries for context continuity

Model Selection

The agent automatically selects the best Qwen3 model for your hardware:

RAM → Model Tier
━━━━━━━━━━━━━━━━
2 GB  → Qwen3-0.6B
4 GB  → Qwen3-1.7B
6 GB  → Qwen3-4B
10 GB → Qwen3-8B
18 GB → Qwen3-14B
36 GB → Qwen3-32B

All models support native tool calling — no prompt injection or regex parsing needed.