Rikugan / webpage
llms.txt690 starsChanged 38 days agoMITView on GitHub
An agentic AI reverse-engineering assistant that runs natively inside IDA Pro and Binary Ninja. Rikugan is a plugin that embeds a full agentic LLM loop — with its own tool framework, skills system, context management, and chat UI — directly inside your disassembler. It is not an MCP client or a wrapper; it runs in-process with access to the host's scripting API. /malware-analysis, /linux-malware, /deobfuscation, /vuln-audit, /driver-analysis, /ctf, /generic-re, /ida-scripting, /binja-scripting, /modify, /smart-patch-ida, /smart-patch-binja Tools from MCP servers appear with…
Files
# Rikugan (六眼)
> An agentic AI reverse-engineering assistant that runs natively inside IDA Pro and Binary Ninja.
Rikugan is a plugin that embeds a full agentic LLM loop — with its own tool framework, skills system, context management, and chat UI — directly inside your disassembler. It is not an MCP client or a wrapper; it runs in-process with access to the host's scripting API.
## Pages
- [Home](index.html): Overview, feature highlights, skills list, and provider support.
- [Documentation](docs.html): Installation, configuration, usage guide, commands reference, skills, MCP setup, custom skills, and full tool listing.
- [Architecture](ARCHITECTURE.html): Internal design — agentic loop, TurnEvent system, tool framework, subagents, mutation tracking, and thread safety.
## Key Facts
- **Hosts**: IDA Pro 9.0+ and Binary Ninja (UI mode). Windows, macOS, Linux.
- **LLM Providers**: Anthropic (Claude), OpenAI, Google Gemini, Ollama (local), any OpenAI-compatible endpoint.
- **Tools**: 50 shared tools + 6 IDA-only microcode tools + 13 Binary Ninja-only IL tools = 69 total.
- **Skills**: 12 built-in skills including malware analysis, deobfuscation, vuln audit, CTF solving, driver analysis, and binary patching. Custom skills supported via SKILL.md files.
- **Agentic loop**: Generator-based turn cycle — User Message → LLM Stream → Tool Execution → Yield TurnEvents → Next Turn.
- **Exploration Mode** (`/modify <goal>`): 4-phase autonomous binary modification: EXPLORE → PLAN → EXECUTE → SAVE, with user approval gates.
- **Plan Mode** (`/plan <msg>`): Generate a numbered plan, get user approval, execute step-by-step.
- **Mutation Undo** (`/undo [N]`): Every rename, comment, type change, or prototype modification is tracked with a reverse operation.
- **Script Approval**: `execute_python` always shows the full code for user approval before running. Scripts are limited to the disassembler's scripting API.
- **MCP Client**: Connects external MCP servers via stdio + JSON-RPC 2.0 with Content-Length framing. Heartbeat monitoring detects dead servers.
- **Persistent Memory**: `RIKUGAN.md` next to the IDB/BNDB persists findings across sessions. First 200 lines injected into system prompt.
- **Multi-Tab Chat**: Multiple independent sessions per file with fork support and auto-save/restore.
- **Context Management**: Auto-compaction at 80% of configured context window threshold.
## Config Defaults
- `max_tokens`: 16384
- `temperature`: 0.2
- `context_window`: 200000
- `max_tool_result_chars`: 8000
## Built-in Skills
/malware-analysis, /linux-malware, /deobfuscation, /vuln-audit, /driver-analysis, /ctf, /generic-re, /ida-scripting, /binja-scripting, /modify, /smart-patch-ida, /smart-patch-binja
## MCP Config Format
```json
{
"mcpServers": {
"server-name": {
"command": "python",
"args": ["-m", "my_mcp_server"],
"env": {},
"enabled": true
}
}
}
```
Tools from MCP servers appear with the prefix `mcp_<server>_<tool>`.
## Installation
Clone the repo, then run the platform installer:
- `./install_ida.sh` (Linux/macOS) or `install_ida.bat` (Windows)
- `./install_binaryninja.sh` (Linux/macOS) or `install_binaryninja.bat` (Windows)
Installers create plugin links, install Python dependencies, and initialize the Rikugan config directory.
## Source
https://github.com/buzzer-re/rikugan
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

