Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
An AI agent can use MCP to drive decompilers like Ghidra and IDA Pro, letting a model enumerate symbols and follow cross-references to understand a feature, but no data exists on how often it gets things right, and both license limits and model hallucination risk mean every answer needs human checking.
Key points
Ghidra MCP provides a plugin and Python bridge so an AI agent can call decompiler functions over HTTP via MCP.
IDA Pro MCP adds cross-reference listing so the agent can follow a function's callers upward.
Models are poor at numeric math and can state false information, so IDA Pro MCP includes an int_convert tool and warns to verify.
The MCP spec requires a human to deny calls, so a client that asks first is recommended.
Decompiling may be banned by licenses; EU Article 6 only allows it when indispensable for interoperability, and US 1201F is narrow.
Techniques
- MCP client-server bridge
- cross-reference tracing
- function and symbol enumeration
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
You want to know how an app signs requests, but you only have compiled code. A decompiler makes it readable, but gives you thousands of functions named like sub 401 A20. Reading them
takes days. An AI agent can read them for you. It opens only the functions it needs and follows the calls between them. How does it reach your decompiler?
Bridges like this usually add two parts. A plugin inside the decompiler answers HTTP requests from your own computer. A bridge program turns the agent's requests into those HTTP requests.
Ghidra MCP has a plugin and a Python bridge like this. The agent talks to the bridge through MCP, the model context protocol. The app running the agent is the MCP client and the bridge is the
server. The client asks for the tool list, the model picks a tool, and the bridge runs it and returns text. The agent can first list methods, classes, imports, and exports as Ghidra MCP's
tools do, then pick one to open. IDA Pro MCP for the IDA decompiler also lists cross-references, places in the code that point to a function or data, so the agent can follow a function's callers
upward. The agent can be wrong. The IDA Pro MCP readme says models are poor at number-based maths, so use its int_convert tool. It also warns models can state false information. Read the
function yourself. Some tools can rename functions, patch bytes, and run Python on your computer. The MCP spec says a human should be able to deny calls, so use a client that asks first. Use this
to understand a feature, not to copy it. Many licenses ban decompiling, so read yours. EU Article 6 allows it only when indispensable for your own program to work with the app. One summary says US
1201F is narrow and doesn't cover general study. If your license and local law allow it, start with a local bridge and a client that asks first. Check where
your model runs because it reads tool results. This isn't legal advice. How often is the agent right? No source measures that, so check each answer.