GhidrAssist

Author: Jason Tang

An advanced LLM-powered plugin for interactive reverse engineering assistance in Ghidra.

Description

GhidrAssist integrates Large Language Models (LLMs) into Ghidra to provide intelligent assistance for binary exploration and reverse engineering. It supports any OpenAI v1-compatible API, including local models (Ollama, LM-Studio, Open-WebUI) and cloud providers (OpenAI, Anthropic, Azure).

Key Features

Core Functionality:

• Code Explanation - Explain functions and instructions in both disassembly and decompiled pseudo-C
  • Security analysis panel showing risk level, activity profile, and API usage
  • Editable summaries with user-edit protection from auto-overwrite
• Interactive Chat - Multi-turn conversational queries with persistent chat history
• Custom Queries - Direct LLM queries with optional context from current function/location

Graph-RAG Knowledge System:

• Semantic Knowledge Graph - Hierarchical representation of binary analysis
  • 5-level semantic hierarchy: Statement → Block → Function → Module → Binary
  • Pre-computed LLM summaries enable fast, LLM-free queries
  • SQLite persistence with JGraphT graph algorithms
  • Full-text search (FTS5) on summaries and security annotations
• Community Detection - Automatic module discovery via Leiden algorithm
  • Groups related functions into logical modules
  • Hierarchical community structure with summaries
  • Visual graph exploration with configurable depth
• Security Feature Extraction - Comprehensive security analysis
  • Network APIs: POSIX sockets, WinSock, DNS, SSL/TLS, WinHTTP, WinINet
  • File I/O APIs: POSIX, Windows, C library functions
  • Crypto APIs: OpenSSL, Windows crypto, platform-specific
  • String patterns: IP addresses, URLs, domains, file paths, registry keys
  • Risk level classification (LOW/MEDIUM/HIGH) and activity profiling
• Semantic Graph Tab - Visual knowledge graph interface
  • Graph view with N-hop depth exploration
  • List view of all indexed functions
  • Semantic search across summaries
  • One-click re-indexing and security analysis

Advanced Capabilities:

• Extended Thinking/Reasoning Control - Adjust LLM reasoning depth for quality vs. speed trade-offs
  • Support for OpenAI o1/o3/o4, Claude with extended thinking, and local reasoning models
  • Configurable effort levels: Low (fast), Medium (balanced), High (thorough)
  • Per-program persistence - different binaries can use different reasoning levels
  • Provider-agnostic implementation (Anthropic, OpenAI, Azure, LiteLLM, LMStudio, Ollama)
• ReAct Agentic Mode - Autonomous investigation using structured reasoning (Think-Act-Observe)
  • LLM proposes investigation steps based on your query
  • Systematic tool execution with progress tracking via todo lists
  • Iteration history preservation showing all investigation steps
  • Final synthesis with comprehensive answer and key findings
  • Accurate metrics (iterations, tool calls, duration)
• MCP Integration - Model Context Protocol client for tool-based analysis
  • Works with GhidrAssistMCP for Ghidra-specific tools
  • Conversational tool calling with automatic function execution
  • Support for SSE (Server-Sent Events) transport
• Function Calling - LLM can autonomously navigate binaries and modify analysis
  • Rename functions and variables
  • Navigate to addresses and cross-references
  • Execute Ghidra commands
• Actions Tab - Propose and apply bulk analysis improvements
  • Security vulnerability detection
  • Code quality analysis
  • Automated refactoring suggestions
• RAG (Retrieval Augmented Generation) - Enhance queries with contextual documents
  • Add custom documentation, exploit notes, architecture references
  • Lucene-based full-text search
  • Context injection into queries
• RLHF Dataset Generation - Collect feedback for model fine-tuning

Architecture

The plugin uses a modular, service-oriented architecture:

Core Services:

• Query Modes: Regular queries, MCP-enhanced queries, or full agentic investigation
• ReAct Orchestrator: Manages autonomous investigation loops with todo tracking and findings accumulation
• Conversational Tool Handler: Manages multi-turn tool calling sessions
• MCPToolManager: Interfaces with external MCP servers for specialized tools

Graph-RAG Backend:

• BinaryKnowledgeGraph: Hybrid SQLite + JGraphT storage for semantic knowledge
• GraphRAGEngine: LLM-free query engine using pre-computed summaries
• SemanticExtractor: LLM-powered function summarization with batch processing
• SecurityFeatureExtractor: Static analysis for network, file I/O, and crypto APIs
• CommunityDetector: Leiden algorithm implementation for module discovery

Data Layer:

• AnalysisDB: SQLite database for chat history, RLHF feedback, and knowledge graphs
• SchemaMigrationRunner: Versioned database migrations for transparent upgrades
• RAGEngine: Lucene-powered document search for custom context injection

UI Components:

• Tab-based interface: Explain, Query, Actions, Semantic Graph, RAG Management, MCP Servers
• Service orchestration via TabController

Future Roadmap:

• Model fine-tuning using collected RLHF dataset
• Additional MCP tool integrations
• Enhanced agentic capabilities, multi-agent collaboration
• Embedding-based similarity search

Screenshots

Screenshot

https://github.com/user-attachments/assets/bd79474a-c82f-4083-b432-96625fef1387

Quickstart

• If necessary, copy the binary release ZIP archive to the Ghidra_Install/Extensions/Ghidra directory.
• Launch Ghidra -> File -> Install Extension -> Enable GhidrAssist.
• Load a binary and launch the CodeBrowser.
• CodeBrowser -> File -> Configure -> Miscellaneous -> Enable GhidrAssist.
• CodeBrowser -> Window -> GhidraAssistPlugin.
• Ensure the RLHF and RAG database paths are appropriate for your environment.
• Point the API host to your preferred API provider and set the API key.
• (Optional) In the Analysis Options tab, set the Reasoning Effort level (None/Low/Medium/High) for models that support extended thinking.
• Open GhidrAssist with the GhidrAssist option in the Windows menu and start exploring.

LLM Setup

GhidrAssist works with any OpenAI v1-compatible API. Setup details are provider-specific - here are some helpful resources:

Local LLM Providers:

• LM Studio - Easy local model hosting with GUI
• Ollama - Command-line local model management
• Open-WebUI - Web interface for local models

Cloud Providers:

• OpenAI API
• Anthropic Claude
• Azure OpenAI

LiteLLM Proxy (Multi-Provider Gateway):

• LiteLLM - Unified API for 100+ LLM providers
• Supports AWS Bedrock, Google Vertex AI, Azure, and many others
• Select "LiteLLM" as provider type in GhidrAssist settings
• Automatic model family detection for proper message formatting

Recommended Models

For Agentic Mode (requires strong reasoning and tool use):

• Cloud: GPT-5.1, Claude Sonnet 4.5
• Local: GPT-OSS, Llama 3.3 70B, DeepSeek-R1 70B, Qwen2.5 72B

Models with Extended Thinking/Reasoning Support:

• OpenAI: o1-preview, o1-mini, o3-mini, o4-mini, gpt-5 (use reasoning_effort parameter)
• Anthropic: Claude Sonnet 4.5, Claude Opus 4.5, Claude Haiku 4.5, Claude Opus 4.1/4, Claude Sonnet 4 (use thinking.budget_tokens parameter)
• Local: openai/gpt-oss-20b via Ollama/LMStudio (supports effort levels)