Ida-pro-mcp

[th3tuga]

Originally Posted by Shub-Nigurrath

Hi all,
The discussion went on for a while, and I saw my name mentioned a few times. To clarify my point, I was saying that AI, before or above everything, is a gigantic tracking tool. It can track what you do with it on an unprecedented scale

So it's better to avoid the risk altogether and switch to a local OSS model, e.g., via Ollama, which is blazing easy, and use Claude or Codex with another model under the hood.

Thank you @Shub-Nigurrath
This is one of the main reasons I prefer using local models for reverse engineering is the critical need for privacy. There is a significant body of legal precedent demonstrating how easily law enforcement can obtain chat and usage histories from cloud providers to deanonymize users and link their digital activities to their real-world identities.

For commercial targets, I think abliterated models are the best long-term option. Both the privacy concerns above and the fact that they waste huge numbers of tokens on ineffective attempts make me feel that local LLMs are the better path forward for reversing.
Could you provide some practical guidance on integrating and utilizing local large language models, such as Qwen 3.8, with IDA Pro and x64dbg? @squareD, and I have spent the last few days asking for a viable solution; while theoretical frameworks are plentiful, we are struggling to find a concrete, functional implementation. We are specifically seeking the technical "how-to" and the practical steps necessary to bridge these tools effectively.

@Fyyre, Thanks for the answer. I intend to explore Grok Build as well, though I will first determine if I can acquire an account anonymously without using a credit card to ensure there is no traceable link to my actual identity. The privacy concerns mentioned by @Shub-Nigurrath are very real.

[Shub-Nigurrath]

simple use IDAAssist if you want it inside IDA, or IDAAssistMCP. It can run with Ollama too. Go there and help yourself

[th3tuga]

Originally Posted by Shub-Nigurrath

simple use IDAAssist if you want it inside IDA, or IDAAssistMCP. It can run with Ollama too. Go there and help yourself

This appears to be what I was looking for. So far, it seems to be performing well, and I will provide an update as I continue using it.

[chants]

Here are 2 more interesting tools though not strictly MCP, are similar in automating typical reversing workflows

ReverserAI (v1.2)
ReverserAI is a research project designed to automate and enhance reverse engineering tasks through the use of locally-hosted large language models (LLMs). Operating entirely offline, this initial release features the automatic suggestion of high-level, semantically meaningful function names derived from decompiler output. ReverserAI is provided as a Binary Ninja plugin; however, its architecture is designed to be extended to other reverse engineering platforms such as IDA and Ghidra.
https://github.com/mrphrazer/reverser_ai

Rev·Deck — AI-Assisted Reverse Engineering with Ghidra
https://github.com/biniamf/ai-reverse-engineering

[wx69wx2023]

IDA Assist is out, more detail here:

https://fixupx.com/in19h/status/2105827954471105013

https://docs.hex-rays.com/ida-9.5/add-ons/assist

available to IDA Home and Pro subscribers (IDA 9.4 sp1+).

[DARKER]

Some real review of mrexodia ida-pro-mcp with IDA MCP? Your own feelings.

[chants]

Is the process for efficient use that first an MCP tool becomes in widespread mainstream usage, then that data trains a next model, then the next generation of models become highly proficient at using that particular MCP tool? Ive suspected that e.g. GPT works better in Codex and Claude works better in Vlaude Code because the harness tools have now been deeply trained into the model. Sure theCP json schema and its tools are sent in the prompt and such but I feel its more clunky and inefficient until enough training data optimizes efficient usage. Im not sure but it seems LLMs require incremental improvements based on practical usage.