Stop building offical MCP server

Don’t waste your time building an official MCP server - just focus on implementing the long-missing CLI version like this one. AI agents will use it smoothly too, benefiting both humans and machines at the same time.

./ida open-database …

./ida execute-python …

./ida reference …

./ida decompile …

Agreed. Actually, idat.exe already works as a makeshift CLI version. Although it’s not very straightforward, it could certainly be further optimized.

The MCP solves the problem of multiple agents accessing the same IDB through the ida-nexus library. This library can also be used by community projects such as ida_buddy, which does pretty much exactly what you want:

idb open foo.exe          # attach GUI or managed idalib, print a triage summary
idb u sub_401000          # disassemble a function (windbg alias for `disas`)
idb dec sub_401000        # decompile (Hex-Rays)
idb db 0x401000 -n 64     # hexdump 64 bytes
idb x CreateFile          # find symbols by pattern  (alias for `names`)
idb xref_to validate_key  # who references this, with instruction context
idb triage sub_401000     # size up a function before reading it
idb ? sub_401000 + 0x10   # evaluate an expression (alias for `eval`)
idb dt GUID               # inspect a type

Alternatively you can tell you agent to use the ida-nexus CLI:

# IDA Domain API reference
uvx ida-nexus reference "decompile function"

# Execute Python against an IDB (command, script, repl)
uvx ida-nexus python tests/crackme03/elf -c 'db.functions.get_all()'

The MCP is a thin wrapper around ida-nexus (~200 lines of actual logic), so you can rest assured no unnecessary time was wasted :wink: The goal of the MCP is that you install it and your agents will be effective at reverse engineering tasks without additional prompting/skills, but if you want to do that yourself you definitely can!

Do we need MCP for commands like cat/ls/mv/…?
No.

MCP is for small, lesser-known tools. The industry-standard state of the art, IDA Pro, doesn’t need MCP at all—all you have to do is implement a better native CLI version and wait for LLMs to learn it. Not to mention, a well-written ida --help is usually enough for AI agents to operate it.

That said, why is MCP bad?

  1. Too many dependencies: companies without Internet access can’t use it, and it may introduce security risks. Without Internet, uv can’t even be installed.

  2. It’s not ready to use for any AI agent — not a smooth experience.

It would be helpful to reread my explanation above to make sure you properly understand it. You do not have to use MCP if you do not want to, you can use ida_buddy or ida-nexus CLIs and ignore MCP completely.

To be clear:

  • You can install uv without internet and use it in air-gapped environments.
  • You do not need internet access to use ida-mcp.
  • You do not need uv to use ida-mcp.
  • Pretty much every agent framework supports MCP.

The dependency tree is very small and can easily be audited:

ida-mcp 20261008.0.1
├── ida-nexus 0.13.4
│   └── ida-domain 0.5.1
│       ├── idapro 0.0.10
│       ├── packaging 26.3
│       └── typing-extensions 4.16.0
├── packaging 26.3
└── zeromcp 1.10.3

That being said, if you have special deployment requirements or need help setting up IDA MCP you are welcome to reach out to the Hex-Rays support so we can assist you!

Sure, there are a million ways to meet my needs—but all I really want is a native CLI version of IDA Pro.

Put simply: stop chasing AI (MCP, Skills, and whatever hype comes next) and let AI chase you.

You can always use idat (e.g. with -Smyscript.py) or idalib/ida-domain via a Python host process.

Thank you, sir.

I’m not trying to solve a problem here — I just wanted to share my taste in software design. Hope it’s worth a thought.

IDA Pro’s so great you don’t need to bother adapting to AI — just let AI chase after you.