IDAssist

AI-Powered Reverse Engineering Plugin for IDA Pro

Author: Jason Tang

Description

IDAssist is an IDA Pro plugin that integrates LLM-powered analysis directly into IDA's interface, providing AI-assisted binary reverse engineering through configurable LLM providers, semantic knowledge graphs, RAG document search, and supports a wide diversity of LLM providers.

Built with Python and PySide6, IDAssist runs as a dockable panel inside IDA Pro 9.0+ and communicates with LLM providers (OpenAI, Anthropic, Ollama, LiteLLM, and more) to analyze functions, suggest renames, answer questions about code, and build a searchable knowledge graph of an entire binary.

Core Features

Function Explanation — Generate detailed natural-language explanations of decompiled functions with automatic security analysis including risk level, activity profile, security flags, and API detection.

Interactive Query Chat — Ask questions about the binary with persistent chat history. Use context macros (#func, #addr, #line, #range) to inject function code, addresses, or disassembly ranges into queries.

Automated Actions — AI-powered rename suggestions for functions, variables, and types. Review proposed changes in a table with confidence scores, then apply selected actions back to the IDB.

Semantic Knowledge Graph — Build and explore a knowledge graph of the binary's functions, call relationships, data flows, and security characteristics. Includes visual graph rendering, semantic search, and community detection.

RAG Document Search — Upload reference documents (.txt, .md, .rst, .pdf) and use them as context during LLM queries. Supports hybrid text+vector search via Whoosh indexing.

SymGraph Integration — Push and pull function names, variable names, types, and graph data to the SymGraph collaborative platform. Includes a multi-step wizard with conflict resolution for pulls.

Settings Management — Configure multiple LLM and MCP providers, manage SymGraph API credentials, customize the system prompt, and set database paths.

Advanced Capabilities

ReAct Agent

The Query tab supports an autonomous ReAct (Reasoning + Acting) agent mode. When enabled, the LLM plans an investigation strategy, executes tools to gather information, reflects on findings, and synthesizes a comprehensive answer — all automatically across multiple reasoning rounds.

Extended Thinking

Configure reasoning effort levels to control how much the LLM "thinks" before responding:

MCP Integration

IDAssist can connect to external MCP servers for tool-augmented LLM interactions where the model can programmatically inspect functions, read disassembly, query cross-references, and modify the IDB during reasoning. Both URL-based MCP servers (SSE or Streamable HTTP) and stdio-based MCP servers started from a local CLI command are supported. IDAssist also provides built-in internal tools for function calling without requiring an external MCP server.

Function Calling

LLM providers with tool-calling support can invoke IDA analysis functions mid-conversation, enabling iterative investigation without manual intervention.

RLHF Feedback

Provide thumbs-up/thumbs-down feedback on explanations and query responses. Feedback is stored locally and can be used to improve prompt engineering and model selection.

Architecture

IDAssist follows an MVC (Model-View-Controller) pattern:

• Views (src/views/) — PySide6 tab widgets that emit signals on user interaction
• Controllers (src/controllers/) — Connect view signals to service calls, manage state
• Services (src/services/) — Business logic, LLM providers, database access, graph analysis
• Internal Tools (src/services/internal_tools.py) — IDA-specific tool definitions for LLM function calling
• Graph Tools (src/services/graphrag/graphrag_tools.py) — Semantic graph read/write tools for LLM interaction

Key design principles:

• All IDA API calls execute on the main thread via execute_on_main_thread()
• LLM responses stream incrementally to the UI
• Local SQLite databases for persistence (no external database required)
• Singleton service registry with thread-safe initialization

Quick Start

1. Install the plugin (recommended — IDA Plugin Manager):

   hcli plugin install idassist
   

   This automatically installs the plugin and its Python dependencies into IDA's environment.

2. Or install manually (from release tarball):

   Download the latest release zip from GitHub Releases and extract it into your IDA plugins directory:

   Linux / macOS:

   unzip IDAssist-*.zip -d ~/.idapro/plugins/
   

   Windows: Extract the zip into %APPDATA%\Hex-Rays\IDA Pro\plugins\.

   Then install dependencies using IDA's bundled Python (not your system Python):

   Linux / macOS:

   /python3/bin/pip3 install -r ~/.idapro/plugins/IDAssist/requirements.txt
   

   Windows:

   "\python3\python.exe" -m pip install -r "%APPDATA%\Hex-Rays\IDA Pro\plugins\IDAssist\requirements.txt"
   

   Replace `` with your IDA Pro installation path (e.g., /opt/idapro-9.0 or C:\Program Files\IDA Pro 9.0).

   Tip: You can also set the IDAUSR environment variable to a custom directory containing a plugins/ subdirectory.

3. Open IDAssist: Launch IDA Pro, open a binary, and press Ctrl+Shift+A (or Edit > Plugins > IDAssist).

4. Configure a provider: Go to the Settings tab, click Add under LLM Providers, and configure your preferred provider.

5. Analyze a function: Navigate to any function, click the Explain tab, and press Explain Function.

For detailed setup instructions, see Getting Started.

LLM Provider Setup

IDAssist supports the following provider types:

Recommended Models

• Open-weight models: GLM, Qwen, DeepSeek, Nemotron, Gemma, and Muse are recommended families for local or self-hosted analysis. Choose an instruction-tuned, coding, or reasoning model supported by your runtime.
• Hosted frontier models: Opus, Sonnet, and GPT are recommended for code analysis and complex investigations.
• Agentic mode and Actions: Choose a model and serving backend with reliable native tool calling and multi-step reasoning. Extended thinking controls depend on the selected model and provider.

These are family names, not literal API identifiers. In Settings, enter the exact model ID or installed tag provided by your endpoint. See the model guidance for selection criteria and official model resources.

Using the Semantic Graph

The Semantic Graph tab provides a knowledge graph of the binary:

1. ReIndex Binary — Extracts function structure, call graph, and cross-references
2. Semantic Analysis — Generates LLM summaries for each function
3. Security Analysis — Detects vulnerability patterns and security-relevant APIs
4. Network Flow — Tracks network operations across the call graph
5. Community Detection — Groups related functions into modules

Explore the graph via the List View (callers, callees, edges, flags), Visual Graph (interactive node diagram with N-hop expansion), or Search (7 query types including semantic search, similar functions, and call context).

Context Menu Actions

Right-click in any Disassembly or Pseudocode view to access:

Requirements

• IDA Pro 9.0+ with Python 3 and PySide6
• Hex-Rays Decompiler (recommended for pseudocode features)
• Python packages listed in requirements.txt

Documentation

• Documentation Index
• Getting Started
• Tab References: Explain | Query | Actions | Semantic Graph | RAG | Settings
• Workflows: Explain | Query | Semantic Graph

Homepage

https://symgraph.ai

License

See LICENSE file for details.