pyobfus-mcp — Model Context Protocol server for pyobfus

pyobfus-mcp exposes pyobfus — the Python obfuscator — to any MCP-capable AI coding agent: Claude Desktop, Claude Code, Cursor, GitHub Copilot, Windsurf, Zed, Codex, CodeBuddy, and anything else that speaks the Model Context Protocol.

Once configured, you can say:

"Check if this FastAPI project is safe to obfuscate, then generate a pyobfus.yaml for it."

and the agent will autonomously call check_obfuscation_risks and generate_pyobfus_config — no copy/paste of CLI commands, no manual config editing.

Tools exposed

All tools return dicts with a status field, a free-text ai_hint, and a machine-readable next_tool field ({tool, reason, args}) so an agent can chain steps deterministically. verify_cmd in protect_project runs an arbitrary command and is therefore opt-in behind PYOBFUS_MCP_ALLOW_VERIFY_CMD=1.

Install

Zero-install (recommended)

If you have uv, no separate install step is needed — point your client at uvx and it fetches pyobfus-mcp (plus pyobfus and the MCP SDK) into an ephemeral environment on first run. No API key required.

{
  "mcpServers": {
    "pyobfus": {
      "command": "uvx",
      "args": ["pyobfus-mcp"]
    }
  }
}

Or install with pip

pip install pyobfus-mcp

This pulls pyobfus and the MCP Python SDK automatically. Then use "command": "pyobfus-mcp" in the per-client configs below.

Configure

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "pyobfus": {
      "command": "pyobfus-mcp"
    }
  }
}

Restart Claude Desktop. The pyobfus tools appear in the tool list.

Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "pyobfus": {
      "command": "pyobfus-mcp"
    }
  }
}

Windsurf

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "pyobfus": {
      "command": "pyobfus-mcp"
    }
  }
}

Zed

In ~/.config/zed/settings.json:

{
  "context_servers": {
    "pyobfus": {
      "command": {
        "path": "pyobfus-mcp",
        "args": []
      }
    }
  }
}

Claude Code

claude mcp add pyobfus pyobfus-mcp

Example session

User:  Can you check whether this Python project is safe to obfuscate?
       Path: /Users/me/code/my-api

Agent: [invokes check_obfuscation_risks("/Users/me/code/my-api")]
       I found 2 high-severity and 3 medium-severity patterns. FastAPI is
       detected, so I'd suggest the `fastapi` preset. Want me to generate
       the config?

User:  Yes please, write it.

Agent: [invokes generate_pyobfus_config("/Users/me/code/my-api",
         preset_override="fastapi", write=True)]
       Wrote pyobfus.yaml. Next: pyobfus /Users/me/code/my-api -o dist/
       -c pyobfus.yaml

Debugging obfuscated code with your AI assistant

The killer feature: keep AI-assisted debugging even after you obfuscate.

User:  Here's a crash from prod. Can you help?
       [pastes traceback full of I0, I1, I2...]

Agent: [invokes unmap_stack_trace(trace, "path/to/mapping.json")]
       Reversed. The crash is in Calculator.add() called from
       main() — 'Calculator' object has no attribute 'add_x'. Looks like
       a typo in the method call site…

Verifying tool integrity

Tool poisoning / "rug-pulls" — a server's tool descriptions silently changing after you've reviewed and trusted them — are the #1 threat in the 2026 MCP security baseline. pyobfus-mcp-verify checks the currently-installed package's tool descriptions/schemas against a manifest frozen at release time:

pyobfus-mcp-verify
# OK: tool descriptions match the shipped manifest (digest 5cf4c902d460...).

A mismatch (exit code 1) means the installed tools differ from what the release shipped — worth investigating before trusting the server. This is a self-consistency digest, not a cryptographic signature: it proves the package matches its own shipped manifest, not that a specific party signed it. For stronger assurance, compare the printed digest against the one published in the GitHub Release notes for the version you installed.

Independent security scan

The published PyPI package (not a local build) has been scanned with Cisco's open-source mcp-scanner — a live initialize → tools/list handshake followed by YARA malware-signature and dependency-vulnerability analysis against all 8 exposed tools. Result as of pyobfus-mcp 0.3.6 (2026-08-20): 8/8 tools SAFE, 0 findings. Full reproducible command, raw output, and an honest statement of what this scan does and doesn't cover: docs/MCP_SECURITY_SCAN.md.

License

Apache-2.0. Same as the main pyobfus package. The pyobfus Pro features remain license-gated; this MCP server only wraps the community-tier tools.

Links

• Main package: https://pypi.org/project/pyobfus/
• Source: https://github.com/zhurong2020/pyobfus
• AI integration strategy: docs/AI_INTEGRATION_STRATEGY.md
• MCP specification: https://modelcontextprotocol.io/