Reverse-engineering MCP for your AI assistants — an MCP server plus a one-click multi-agent installer.

English · 简体中文

v1.0.1 — install the right package. The published CLI is metis-reverse-cli (npm dist-tag latest → 1.0.1). The older npm name metis-reverse is still stuck at 0.0.1 and is not this release. Prefer npx metis-reverse-cli@latest setup. The engine repo also has a moving Git tag latest (currently = v1.0.1) for uvx / source downloads.

1. Install & connect. You need Node.js 18+ and uv (the MCP engine is pulled with uvx from GitHub; there is no PyPI package yet).

Step-by-step beginner guide: docs/INSTALL.md.

npx metis-reverse-cli@latest setup

Or install the CLI globally, then run the same wizard:

npm install -g metis-reverse-cli@latest
metis setup          # same binary as `metis-reverse`

setup finds coding agents on your machine, shows what it will write (MCP + Skill + Cursor rules), keeps a .bak of existing MCP configs, and asks before writing. When you see Setup complete! Please restart your selected agents to activate Metis Reverse., restart the agent. Check with:

metis doctor
# or: npx metis-reverse-cli@latest doctor

Remove only Metis Reverse (other MCP servers stay):

metis uninstall

Dry-run / preview without writing:

metis preview
# same as: metis setup --dry-run

{
  "mcpServers": {
    "metis-reverse": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/Wholiver/metis-reverse.git@latest",
        "metis-reverse"
      ]
    }
  }
}

git clone --branch latest --depth 1 https://github.com/Wholiver/metis-reverse.git
cd metis-reverse
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"

2. Try it. Build the bundled demo, then ask your agent to inspect it with the re_* / js_* tools.

cd demo && make

Inspect demo/demo_target. Triage the binary, decompile the interesting functions, pull evidence for names and types, then verify any hypothesis before you trust it.

Evidence stays attached. re_query_evidence and re_add_finding keep provenance on symbols and addresses so the agent can cite why a name or type is trusted.	Inspect first, then go deeper. re_inspect / re_triage / re_capabilities give a cheap first pass; optional backends (e.g. Ghidra via GHIDRA_INSTALL_DIR) are used when present, otherwise the built-in path still runs.
Decompile, refine, transfer. Task-level tools like re_decompile_function, re_refine_model, and re_diff_and_transfer move analysis forward without dumping raw one-shot APIs on the agent.	Probe when static is not enough. re_emulate_slice and re_run_probe_script pull concrete values from a slice or a probe and feed them back into the session.
Session state is shared. re_session_manager, artifact lists, and findings keep later steps on the same tree — the agent does not start from zero every turn.	Verify before you ship a claim. re_verify_hypothesis and related checks push the agent to test ideas instead of narrating them.

About 29 MCP tools in the v1.0.1 engine (handlers under re_* and js_*). Grouped by task below.

Bins from the npm package: metis and metis-reverse (same entrypoint).

pytest tests/unit/
pytest tests/e2e/
pytest --cov=metis_reverse tests/

Metis QQ group: 801193315

This project supports the LINUX DO community.

Built with (and optional external tools when you install them yourself):

_{Released under the MIT License. Engine + docs: github.com/Wholiver/metis-reverse. CLI on npm: metis-reverse-cli.}