# DecompAI

> Gradio chat application whose LangGraph ReAct agent analyses an uploaded x86 Linux ELF binary: it lists and disassembles functions, decompiles them with Ghidra, drives radare2 and gdb, and runs shell commands and Python scripts in a privileged Kali Linux runner container. Workspaces are keyed by the binary’s SHA-256, so uploading the same file again restores the session. It came out of a school research project, and its CTF results are the authors’ own report.

- Type: Integration
- Targets: Binary
- Category: Independent agent workflows
- Relationship: Independent agent integration
- Sources reviewed: 2026-10-19
- Added to the catalog: 2026-10-19
- Original resource: <https://github.com/louisgthier/decompai>
- Primary source: <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/README.md>
- Catalog page: <https://awesomerea.org/resources/decompai/>
- Record JSON: <https://awesomerea.org/resources/decompai/record.json>
- Chinese version: [简体中文](https://awesomerea.org/zh/resources/decompai/index.md)

## Requirements & limits

- The reviewed main commit of 2025-11-10 containerised the app. The repository has no tagged releases; the README points to app and runner images tagged 1.0.0 on Docker Hub. Only x86 Linux ELF binaries are supported, and QEMU-based ARM, MIPS and Windows PE support is listed as future work.
- The app needs Python with the pinned requirements.txt (Gradio 5.25, LangGraph and LangChain) and Docker. docker compose builds or pulls a Kali runner image with kali-linux-headless, Ghidra 11.3.1, OpenJDK 21 and radare2 plugins, and the app controls the host Docker daemon through a mounted /var/run/docker.sock. Runner containers start with --privileged on Docker’s default network, and the Gradio interface listens on 0.0.0.0:7860 by default; the README asks maintainers to audit any added dependencies.
- An OpenAI or Gemini API key and an LLM_MODEL setting go in .env. The README suggests gemini-2.5-pro or gemini-2.5-flash, while the code defaults to gpt-4o-mini; disassembly, decompiled code and tool output are sent to that provider. Users upload their own binaries. None are committed, and tests/test_disassemble.py expects a binaries/ch2.bin that the repository does not include.
- The README reports validation on Root-Me CTF challenges, consistently solving difficulty-3-of-5 exercises automatically or semi-automatically, with no published runs or logs. Session data under /tmp/decompai_analysis_sessions by default is lost on reboot.
- LICENSE is MIT text with an added clause requiring any use, modification or distribution to display a “Based on DecompAI” attribution, so GitHub does not recognise it as a standard licence; Ghidra, the Kali packages and model services keep their own terms. The README limits use to lawful reverse engineering, education and security research.

## Primary sources

1. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/README.md>
2. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/LICENSE>
3. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/compose.yaml>
4. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/Dockerfile.runner>
5. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/requirements.txt>
6. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/src/main.py>
7. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/src/config.py>
8. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/src/tools/tools.py>
9. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/src/tools/sandboxed_shell/dockerized_bash.py>
10. <https://github.com/louisgthier/decompai/blob/0c2398c09e318be1bb6f6a52d13e572fac7c519b/tests/test_disassemble.py>

## More in Agents and MCP

- [Kong](https://awesomerea.org/resources/kong-agentic-reverse-engineer/index.md) · Integration
- [Co-op Anything](https://awesomerea.org/resources/coop-anything/index.md) · Integration
- [DeepExtract (DeepExtractIDA)](https://awesomerea.org/resources/deepextract-ida/index.md) · Integration
- [Mizuchi](https://awesomerea.org/resources/mizuchi-matching-decompilation-pipeline/index.md) · Integration
- [ReAgent](https://awesomerea.org/resources/reagent-cpp-reconstruction-agent/index.md) · Integration
- [lisa.py](https://awesomerea.org/resources/lisa-lldb-mcp/index.md) · Integration

## About this entry

The description was reviewed against primary sources.
Original editorial content: MIT. Linked projects, papers and assets retain their own rights and licenses.