🤖 IDA Copilot

An AI assistant for IDA Pro 9.x — chat with an LLM inside IDA and let it inspect, decompile, rename, retype and comment your binaries for you.

Powered by Pydantic AI · OpenAI-compatible endpoints · Qt chat UI

English · 中文

---

✨ Features

• 🪟 Dockable chat window — opens with Ctrl+Shift+C or Edit > Plugins > IDA Copilot, docks to the right of the workspace.
• 🧠 Pydantic AI agent — a real agent loop with streaming, tool calling and thinking support.
• 🔧 30+ IDA tools — the agent can read and modify the database: pseudocode, disassembly, xrefs, strings, segments, renaming, retyping, structs/enums, comments, and more.
• ⚡ Streaming output — tokens appear as they are generated, rendered as Markdown (tables, code blocks, lists…).
• 💭 Thinking & tool-call blocks — collapsible cards, collapsed by default, expandable anytime.
• 🛠️ run_idapython — the model can execute IDAPython snippets against the live database (sandboxed: IDA modules + Python stdlib only).
• 🔌 Any OpenAI-compatible endpoint — OpenAI, DeepSeek, Qwen/DashScope, Ollama, vLLM, etc.
• ⏱️ Configurable timeouts — request and tool-execution timeouts.
• 🧵 Non-blocking — the agent runs on a background worker thread; IDA's UI never freezes.

🖼️ Screenshots

📋 Requirements

• IDA Pro 9.x (GUI, 64-bit)
• IDA's embedded Python 3.12 (or a Python 3.10+ interpreter selected via idapyswitch)
• Python packages: see requirements.txt

🚀 Quick Start

# 1) Install the Python dependencies into the interpreter IDA uses
python -m pip install -r requirements.txt

# 2) Copy / symlink this folder into your IDA plugins directory
#    Windows:  %APPDATA%\Hex-Rays\IDA Pro\plugins\ida_copilot
#    macOS/Linux: ~/.idapro/plugins/ida_copilot
#    (or drop it into IDA's own <ida>/plugins/ directory)

# 3) Restart IDA

🎮 Usage

1. Open the chat window: press Ctrl+Shift+C (or Edit > Plugins > IDA Copilot).

2. Click ⚙ Setting, fill in your Endpoint, API Key and Model, then Save.

3. Ask away — for example:

   What does sub_401000 do? Rename its local variables to something meaningful and add comments at each call site.

The agent will stream its answer and, where useful, call IDA tools that you can expand to inspect.

⚙️ Configuration

Settings are persisted to a local JSON file:

Field	Description
Endpoint	OpenAI-compatible base URL, e.g. https://api.openai.com/v1
API Key	Secret key (stored locally, never sent anywhere else)
Model	Model name, e.g. gpt-4o, deepseek-chat, qwen-plus
Max Context	Soft cap on input tokens
Max Output	max_tokens for completions
Request Timeout	Timeout (s) for model API requests
Tool Timeout	Timeout (s) for tool execution
Thinking	Enable reasoning output where supported
System Prompt	Optional custom system prompt

Config file locations:

• Windows: %APPDATA%\Hex-Rays\IDA Pro\ida_copilot.json
• macOS / Linux: ~/.idapro/ida_copilot.json

Environment variables are used as defaults: OPENAI_BASE_URL, OPENAI_API_KEY, OPENAI_MODEL.

🧰 Agent Tools

The agent can call any of these tools:

Read / inspect

get_ea_by_name · get_name_at_ea · get_current_ea · get_function_list · get_segments · get_strings · get_xrefs_to · get_xrefs_from · get_function_info · get_disassembly · get_function_disassembly · get_pseudocode · get_data_info · get_type_info · get_comments · list_local_types

Write / edit

set_name · set_comment · set_function_comment · set_function_prototype · set_type_at_ea · set_variable_name · set_variable_type · create_struct · create_enum · add_struct_member · rename_struct_member · del_struct_member · apply_struct_type

Execute

run_idapython — runs IDAPython code against the database (sandboxed).

🛡️ Safety

• run_idapython only allows IDA modules + Python stdlib; host modules (os, sys, subprocess, socket, pathlib, ctypes, requests, …) are blocked so the model cannot escape to the host OS.
• Timeouts protect against runaway requests and long-running tool calls.

❓ FAQ

Which models work?

Any OpenAI-compatible chat model: OpenAI, DeepSeek, Qwen/DashScope, Ollama (local), vLLM, Groq, Mistral, etc. Point the Endpoint at its /v1 base URL.

Does it need Hex-Rays / the decompiler?

No. Pseudocode tools return a clear message when the decompiler is unavailable; all other tools still work.

Is my API key stored safely?

It is stored in plaintext in a local JSON config file (never sent anywhere other than your configured endpoint). Treat the file like you would any local secret.

📁 Project Layout

ida_copilot/
├── ida-plugin.json      ← IDA 9.x directory-plugin manifest
├── requirements.txt     ← Python dependencies
├── main.py              ← plugin entry point (hotkey, menu, docking)
└── ida_copilot/
    ├── config.py        ← settings model + JSON persistence
    ├── tools.py         ← IDA tools exposed to the agent
    ├── agent.py         ← Pydantic AI agent + streaming worker
    └── ui.py            ← Qt chat window + settings dialog

🤝 Contributing

Contributions, issues and feature requests are welcome. Feel free to open a pull request or an issue.

📄 License

Distributed under the Apache License 2.0.

---

中文文档