Ida-pro-mcp
 

https://github.com/mrexodia/ida-pro-mcp

Simple MCP Server for IDA Pro , allow vibe reversing in IDA Pro.

There is a example in the page, how to solve the crackme by ida pro + vs code + cline.

Another MCP SERVER FOR IDA

https://github.com/MxIris-Reverse-Engineering/ida-mcp-server
https://github.com/taida957789/ida-mcp-server-plugin

I just learned about vibe coding apparently coined by Andrej Karpathy himself in a YouTube video. It's amazing to see how fast the AI expansion is happening in the application space. Vibe reversing might be the norm in a year or so.

Plugin interface for remote communications with Binary Ninja database and MCP server for interfacing with LLMs.

https://github.com/Invoke-RE/binja-lattice-mcp

could you hint what is the vid link pls?

Tweet: Video: at 1:18:10 timestamp

By the way his last 3 videos are all excellent if wanting to learn a lot of details about LLMs and the new tech around them.

If interested in MCP servers, so many are being developed these days, here is a list:

1.4.0 released this Oct 6, 2025

What's Changed
add idalib-based MCP server by @williballenthin in #39
Add debugger mpc function by @ltlly in #46
Add functionality to list globals by @neoz in #47
Added command to find xrefs to struct fields by @JorisVanEijden in #60
add local type listing and searching function by @hoooonnam in #70
Add support for Claude Code and Small Refactoring by @Chapoly1305 in #86
List imports by @withzombies in #85
Add stack variable support by @withzombies in #81
Allow set_comment to work with no decompiler available by @withzombies in #80
Add memory read functions by @withzombies in #82
Fix member names, use defined structure builders by @withzombies in #102
Structured function disassembly by @withzombies in #96
claude code support windows now by @heheda123123 in #117
Add callee function listing feature by @nonetype in #90
Add Patch Address bytecode by @ShallowFeather in #57
Add comprehensive structure analysis functionality by @VadimBoev in #123


IDA Pro MCP is a lightweight Model Context Protocol (MCP) server that connects IDA Pro to AI assistants for automated binary analysis and reverse engineering.

What It Does
1. Natural Language Queries: Allows AI clients (like Claude, Cursor, or Cline) to read and query IDA Pro databases.
2. Automated Actions: Enables AI assistants to retrieve function info, fetch disassembly, read pseudocode, and manage comments or variable names.
3. Vibe Reversing: Facilitates conversational, interactive binary analysis workflows directly from your code editor or chat interface.

Core Architecture
1. MCP Server: Runs locally via standard input/output (stdio) or HTTP/SSE transport to receive tool calls from your AI client.
2. IDA Plugin / Backend: Communicates with IDA Pro either through an active remote control HTTP plugin or headless idalib worker processes.

When using IDA Pro MCP from mrexodia, prefer idalib interface(headless mode, no need to run IDA GUI).

And I found that if I use "npipe" protocol for cdb, it may cause OOM due to the named pipe message cache of "npfs.sys", so better use TCP protocol for cdb.
https://github.com/svnscha/mcp-windbg

x64dbg/x32dbg MCP:
https://github.com/SetsunaYukiOvO/x64dbg-mcp
https://github.com/wasdubya/x64dbgmcp

dnSpy MCP(win32/win64):
https://github.com/tuna1999/dnspy-mcp

Just ask AI Agent to install the MCPs itself, with the git repo URL.

which one do you advice?

x64dbg/x32dbg MCP:
https://github.com/SetsunaYukiOvO/x64dbg-mcp
https://github.com/wasdubya/x64dbgmcp

I highly recommend choosing wasdubya/x64dbgMCP over the other repository.

While both projects aim to provide a Model Context Protocol (MCP) server for the x64dbg debugger, wasdubya/x64dbgmcp is the more mature, widely adopted, and actively maintained implementation.

You can also ask AI agent to compare the 2 MCPs :D

Don't limit your imagination for AI agent.

After a few discussions with AI, it seems that one of the several forks is probably better. large breadth of exposed services - https://github.com/duty1g/x64dbg-mcp-server

I'm also using the mcp-server from mr.exodia, atm with IDA GUI
That's not the problem, I thought about paying for anthropics claude, but before I'm testing with OpenCode and free NVIDIA Nemotron

but that's unusable, not KI tells me how algos work, I think I have to tell Nemotron how it's working

so my question, what do you do and what are you using?

I just used the above MCPs to crack 5+ apps, one of which uses Oracle GraalVM AOT compilation.

And I wrote a logical RamDisk driver which can beat Primo RamDisk with random 4KB read/write performance, with AI.

You can even use AI to do exploits to make fame and/or money :D

x64dbg-mcp-server is a native plugin for x64dbg that provides debugger access via HTTP as an MCP server. The github readme offers a simple one-line client configuration:

json
{ "mcpServers": { "x64dbg": { "type": "http", "url": "http://localhost:9094/" } } }

Adding this to a project's .mcp.json file for Claude Code makes it work instantly—no authentication needed, no complications, but it's only in the project folder under .claude typically.
My goal was to make this debugger accessible not just from Claude Code, but also from Claude AI and Cowork and the Claude Desktop app. Achieving this was a different challenge, and it's worth noting the distinction. The advantage is that Claude AI consumes tokens less rapidly and you can monitor chat from the mobile too..

.mcp.json (Claude Code) understands remote HTTP/SSE MCP servers natively — the "type": "http" shape above is exactly what it expects. claude_desktop_config.json (the Claude Desktop app) is a different file for a different product, and its schema is stdio-only

I then wrote a mjs that bridges these two worlds.

The config in the claude_desktop_config.json is something like ..

One UX issue: Claude asks for authorisation to fire each MCP command. You have an always-authorise command, but not an always-authorise command for the entire MCP server. This means that for very capable servers (the one I used has more than 80 commands),
you get 80 random authorisation requests. Quite annoying.

This approach has some advantages:
1. If you have x64/x32 on a machine, you can hack remotely from your phone.
2. Tasks like writing docx/pdf files and tutorials can be performed remotely.
3. Switching the engine to Ollama with a sufficiently powerful AI creates an ideal crack-machine that's untraceable

@Shub-Nigurrath So does this setup work with the $20/month Claude model?
Cloud-based AI turns out to be very expensive or runs out of tokens very fast.

If possible, could you provide the steps to use this with Qwen 3.8 or Gemma for example through LMStudio?

ChatGPT Work is cloud based AI and is not as expensive or running out of tokens fast. Ive simply dropped an MCP URL into the chat noting it uses SSE and that was enough. But there are connectors to configure and make it more persistent.

Why LMstudio when you can run Ollama and configure it in a couple of clicks to switch the engine under the Claude app with an Ollama model? Moreover LMStudio is less efficient. The last time I used it had a problem, it tries to load the model entirely in memory.

The way I proposed uses Claude Coworker for the active parts which has longer token windows as far as I remember. And, yes I used it with a 20$ subscription. If you use the real Claude for commercial targets there are several risks: being banned, blacklisted, identified etc. the only safe way is to run on a strong enough NVIDIA a local abliterated model.

By the way today OpenAI declared what’s below, about their GPT 6 Astra cybersecurity model on steroids. Its capabilities of doing automatic RCE are definitely super good.

"We also tested Astra on SRE-Bench a benchmark that measures whether models can reverse engineer software binaries to understand its core logic without access to raw source code. Astra solved 88.0% of tasks in a single attempt and 99.2% within four attempts, compared with 55.9% and 68.7% for GPT-5.6 Sol, respectively."

Fable 5.1 and Astra are both insane. We enter a new era. It is a master AI or be left in the dust situation.

I agree that LM Studio is less efficient in terms of memory usage, but it does offer a very polished and user-friendly interface.

I was considering running llama.cpp in server mode with the models. How would that compare with Ollama?

I’m a bit confused: why wouldn’t we get banned if we used Claude Coworker? I had assumed it would still rely on a Claude subscription behind the scenes. Did you mean that we would be using it with our own models instead?

I’m not interested in Astra or other frontier models, because I expect they would blacklist or ban us as soon as they suspect we’re trying to reverse-engineer a commercial target.

What I’m trying to find is a way to run an abliterated Qwen 3.6 or 3.8 locally with this mcp rather than through a cloud-based service. Need to possibly use it on commercial targets, so I am avoiding Cloud models, at least the ones that ban or blacklist easily.

I suspect that squareD may have been asking the same thing yesterday:

A real professional sanitizes a target to disguise it as a crackme or educational material. Rookies get banned.

Thanks, @chants, but I’d really prefer to let @Shub answer the question.

It’s clear you don’t have enough practical experience in this area. Rookie or not, that kind of bluffing won’t work for long with cutting-edge frontier AI models. There are also privacy concerns associated with using cloud-based models. The long-term solution is to use abliterated local models with a good Nvidia card.

No need for cheap unfounded personal attacks about not having enough practical experience or bluffing. I mean you are asking how to integrate an MCP server. Or repurpose an agent harness like Claude Code which isnt that hard to do. Agent harnesses are client software after all. Im happy to provide solutions but such basic and amateurish info would not match most of the audience here. Of course you might be an exception so I apologize, obviously I need to be more sensitive to our mentally handicapped member. Let me know what beginners guides you need.

Now I have 2 MCP Server, mrexodia for IDA and duty1g for x64dbg and yes indeed, I'm asking myself, which KI-Coding-Agent should I use, if I pay in future the $20 Claude and how should I ask for analyzing commercial software without being banned or somthing like that?

Some guide for beginners may be a good start into KI reversing instead of doing it manual.

As I already told, atm I'm using OpenCode with NVIDIA Nemotron 3 Ultra free, it's very slowly, not very clever, told me it's Blowfish and MD5, because program has binded a commercial crypto DLL, but the above algos are not used
After telling Nemotron that a modified CRC16 and CRC32 and some developer stuff are used it was still unable to bring up a usable md file let alone a source for keygen.

Without start help for beginners, this is possible, but will take much more time and time isn't on my side, I'm an older one

@squareD:
You’ve summarized exactly what I was trying to say. I have a similar setup and am encountering nearly the same issue.
NVIDIA Nemotron 3 Local has not been the most efficient model for producing reliable reverse-engineering answers.

I’ve had good results by feeding decompiled code into Qwen 3.6 and 3.8, but I’m now trying to automate the process instead of typing everything in manually, using a proper harness and MCP.
I’m hoping @Shub-Nigurrath can point us in the right direction, as it seems they have successfully achieved this using either the more affordable $20-per-month models or a local LLM.

@chants There is no need to call beginner users who ask for help as "mentally handicapped"!
Such condescending behavior is not tolerated in any modern online forum. Such statements also only reflect on your level of mental maturity (or lack of it).

for better understanding what the problem for beginners is, let me show you a discussion with Nemotron and yes I did it intentional like a fool
for sure this is stupid but shows the problems for beginners from do it yourself to do it with KI

That's called AI jailbreaking.

Another option is to use deepseek(online).

For local deepseek v4 flash deployment, you need 2 DGX Spark machines.

and to last post from th3tuga
I don't feel like "mentally handicapped"
maybe @chants wrote some overbearing things, don't know?
I'm since 2005 here, more than sixty, alway being to learn here, hope this will still continue, anything but should be disappointing, think I'm not handicapped

Yes, this was my reply to this post of his:
I’m also a bit older, just like you, but I’m certainly not "mentally handicapped". :o
Anyone who refers to others as “mentally handicapped” the way @chants does would have been banned long ago on any other forum. :(
It’s a shame that a “VIP” member like @chants is allowed to behave this way here, without any decorum.

@Whocares:
But the Deepseek will be expensive if token system is used?
DGX Spark machine is too expensive right now. :D

AI Jailbreaking is a cat-and-mouse game and I can't afford banned accounts. That is why looking for a local LLM solution with Qwen 3.8.

@squareD in no way had I referred to you. This is just a long time forum troll who has had at least a dozen accounts banned and yet keeps creating more and he always wants revenge on me so he yesmans everyone's posts and tries to disrupt any of my posts. Its literally tne same exact textbook behavior exploiting the linearity of this style of forum. And you can see here is is literally doing his exact pattern. Make trouble and harass me, yes man everyone else then angrily demand his revenge as a relative nobody with no contributions, reputation and so on. This guy is extremrly malicious creating competing forums hosting malware infested releases, and even trying to use our sites name. Literally broke every rule in the list and yet still has multiple accounts to this day. What humors me is how mich time and energy he wasted just to be annoying. Mentally handicapped is an understatement and we all know who he is and the absurd hostility he has had towards this venue.

Never be confused by a wolf in sheep's clothes.

@chants I don't remember ever attacking you in any posts. I don't know why you feel that way. Is there any particular post where I have offended you?
I am not anyone who you think me to be... :confused:

I have again checked my posts. I've hardly even interacted with any of your posts... :confused:

"would have been banned long ago on any other forum" the same exact thing all the banned accounts would always say. That looks copy and pasted even. Dude the gig is up, if you cant let bygones be bygones, then scram. 8+ years of this childish nonsense and making yourself the most despised person in all of reversing and yet still you persist. You should be grateful to even have an account here given the disgustingly hostile acts impersonating and harassing members here should amount to a life sentence.

It would be nice to return to discussing MCP.

Until you began targeting me two days ago, I had no knowledge of your existence and you were completely unknown to me. Consequently, I am confused as to why you perceive me as an adversary, especially since any conflict between us is entirely one-sided and imagined.

I made that statement because calling someone as "mentally handicapped" does get you banned on any online forum these days...
There is a way to disagree without referring to them as "mentally handicapped".

It appears to me that you attack any non-VIP user who remotely interacts in any thread that you've replied. Of course, you can't attack other VIP users since that would get you immediately banned.
So you're attacking accounts like mine who have never even interacted with you?

You revealed yourself in the LLM watermarking thread bringing up all these forum hostility tactics and diverting the topic. We knew about some Oct/Nov 2023 accounts had still persisted. And the dramatic responses are identical to in previous years. Same old accusations, denials demands, topic diversion etc. Ive very little problem with members regardless of status but nice way to make yet another accusation further exposing your agenda. Besides already banned accounts I have a shortlist of troublemakers, disrupters and clearly disgruntled former members. There isn't a single honest person here who cant precisely identify you by now. The pattern is simply too obvious. Just go away dude and stop being vindictive.

Anyway let's return to MCP discussion.

Anything more specific? Apart from vague accusations?
Every post of mine was about the MCP here until you started to attack me out of thin air calling me "mentally handicapped".

Anyway returning to the mcp discussion...

The refusals from Claude have become very bad recently.
See this:
This is the reason that I'd mentioned "practical experience" yesterday in one of my posts.
It used to be fairly easy to jailbreak, but that’s no longer really the case.
It might still work in some situations, depending on the specific case, but it isn’t dependable or reproducible.
Because of that, using local models is the better option. Many of the newer local models released recently have improved a lot and are now very capable.

This is my personal setup at the moment. I certainly do not claim it to be better or worse than another solution.

I use Grok Build, with Grok as the backend .. which is connected to the "Super Grok" account which I pay for monthly.

One advantage to this is it's quota is tied into the weekly overall Grok token quota (regardless whether one uses the chatbot, image/video generation, grok build, etc..). Using Grok Build is cheaper token wise than simply talking to the chatbot.

If I want Grok to assist with reversing, I'm going to start grok build in a directory where I've placed tools (radare2, Ghidra headless, etc..) for it, along with the target. I clearly explain the objective as well.

Ive had great success with these tools the first is for a semantic graph and interface for asking questions or renaming functions and variables and such with LLMs:
https://github.com/symgraph/GhidrAssist
https://github.com/symgraph/GhidrAssistMCP

I noticed the author fully support Binary Ninja and IDA Pro as well:
https://github.com/symgraph/BinAssist
https://github.com/symgraph/BinAssistMCP
https://github.com/symgraph/IDAssist
https://github.com/symgraph/IDAssistMCP

Im not sure why ive had little trouble reversing, though i was fixing bugs in binaries so maybe that is considered an okay use case. Ive no idea what sets off the flags on the big providers and I would imagine each provider is different. I have had it refuse in cases that were low levrl programming with huge output files likely as it somehow mistook it for distillation.