An AI Research Assistant That Actually Reads Assembly

What it does

This plugin adds a dockable panel to IDA Pro 9.0+ that pipes decompiled functions and disassembly into large language models from OpenAI, Anthropic, Ollama, and others. It streams natural-language explanations, security risk assessments, and rename suggestions back into the interface, while maintaining a local SQLite knowledge graph of call relationships and data flows for semantic search.

The interesting bit

The tool treats the binary itself as a living database: a ReAct agent can autonomously plan investigations, invoke IDA analysis functions mid-conversation via tool calling, and reflect across multiple reasoning rounds without manual prompting. It also pushes and pulls symbol data to the SymGraph collaborative platform, which turns solitary reversing into a team sport with conflict resolution.

Key highlights

• Supports a wide range of LLM providers (OpenAI, Anthropic, Ollama, LiteLLM, AWS Bedrock) with configurable reasoning effort levels from fast queries to deep ~25K-token analysis.
• Context macros (#func, #addr, #line, #range) let you inject precise disassembly snippets into chat queries.
• RAG document search indexes your own .txt, .md, .rst, and .pdf references via Whoosh for hybrid text+vector retrieval.
• Built-in semantic knowledge graph with community detection, N-hop visual graph expansion, and seven query types including similar-function search.
• Local SQLite storage and RLHF feedback collection keep your data and preferences on-disk, not in the cloud.

Caveats

• Requires IDA Pro 9.0+; the Hex-Rays Decompiler is only recommended, but pseudocode-centric features will be hobbled without it.
• Manual setup involves wrestling with IDA’s bundled Python environment instead of your system Python, which the README explicitly warns about.

Verdict

Worth a look if you live in IDA Pro and want an LLM that can call back into the database to resolve cross-references during chat. Skip it if you are not running IDA 9.0+, or if the idea of an autonomous agent poking at your IDB feels too clever by half.