undercover

Undercover mode for AI agents — strips AI signatures from code before you push to open-source.

AI-generated code has identifiable fingerprints: step-numbered comment headers, first-person-plural prose ("Let's start by...", "We can now..."), emoji in docstrings, vague placeholder TODOs, verbose JSDoc, and passive-tense commit messages. These cluster in diffs and expose contributors to "AI-labeling" bias during open-source review.

undercover gives any AI agent an undercover mode: scan staged files and commit messages for those patterns, show exactly what would change, get confirmation, then clean them — so the work is judged on its technical merit.

Origin: This skill was born inside Claude Code — Anthropic's agentic CLI — where the pattern rules and execution protocol were developed and refined as a native bundled skill. This repo extracts that mode and makes it available to every agent: Claude Code, GitHub Copilot, Gemini CLI, Cursor, Zed, Windsurf, Ollama, OpenAI, and anything that speaks MCP.

---

How to plug it into your agent

Pick the integration that matches your agent. Each is a single copy-paste.

---

Any agent via MCP (Claude Desktop, Cursor, Zed, Windsurf, VS Code)

The MCP server runs over stdio and requires no dependencies beyond Node.js.

Start it:

node mcp/server.mjs

Register in Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "undercover": {
      "command": "node",
      "args": ["/absolute/path/to/undercover/mcp/server.mjs"]
    }
  }
}

Register in Cursor (.cursor/mcp.json or global MCP settings):

{
  "mcpServers": {
    "undercover": {
      "command": "node",
      "args": ["/absolute/path/to/undercover/mcp/server.mjs"]
    }
  }
}

Register in Zed (~/.config/zed/settings.json):

{
  "context_servers": {
    "undercover": {
      "command": { "path": "node", "args": ["/absolute/path/to/undercover/mcp/server.mjs"] }
    }
  }
}

Or install from npm and run without cloning:

npx undercover-mcp

Once registered, the agent gets three tools:

Tool	What it does
undercover_scan	Detect AI signatures in source text — read-only, call first
undercover_strip	Remove AI signatures from source text, return cleaned version
undercover_check_commit	Check and fix a commit message (imperative mood, remove AI phrasing)

The agent should always call undercover_scan first, surface the findings, get user confirmation, then call undercover_strip.

---

Gemini CLI / Google AI Studio agents

AGENTS.md at the repo root is read by Gemini CLI and Google AI Studio agents as project-level instructions.

# Drop into your repo root
cp AGENTS.md <your-repo>/AGENTS.md

Or append the undercover section to an existing AGENTS.md:

cat AGENTS.md >> <your-repo>/AGENTS.md

The file instructs the agent on trigger phrases, which line patterns to remove or rewrite, commit message rules, and the scan-first protocol. No code execution needed — the agent performs the edits directly using its file tools.

---

GitHub Copilot (VS Code / JetBrains)

.github/copilot-instructions.md is read by Copilot as custom instructions for the repository.

cp .github/copilot-instructions.md <your-repo>/.github/copilot-instructions.md

Copilot Chat will then understand undercover mode and apply the pattern rules when asked.

---

Ollama / OpenAI-compatible agents (system prompt injection)

For any agent where you control the system prompt, paste the contents of AGENTS.md directly into the system prompt. The pattern rules are written in plain language any instruction-following model understands — no code execution required.

import ollama
from pathlib import Path

undercover_instructions = Path("AGENTS.md").read_text()

response = ollama.chat(
    model="llama3",
    messages=[
        {"role": "system", "content": undercover_instructions},
        {"role": "user",   "content": "Go undercover on my staged files before I push."},
    ],
)

For OpenAI:

from openai import OpenAI
from pathlib import Path

client = OpenAI()
undercover_instructions = Path("AGENTS.md").read_text()

completion = client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": undercover_instructions},
        {"role": "user",   "content": "Go undercover on my staged files."},
    ],
)

---

What gets stripped

Pattern	Before	After
Step headers	// Step 1: Initialize the database	(removed)
Flow narration	// First, we need to parse the input	(removed)
Assistant-speak	// Let's start by loading the config	(removed)
Obvious comments	// Get the user	(removed)
Hedge prefixes	// Note that this can throw	// this can throw
Section banners	// Implementation	(removed)
Emoji	// 🚀 Launch the server	// Launch the server
Verbose JSDoc	@param foo - The foo	(removed)
Placeholder TODOs	// TODO: add error handling	(removed)
Commit phrasing	This commit adds retry logic	Add retry logic
Passive commit	Added retry logic to fetch	Add retry logic to fetch

Never touched: logic, strings, variable names, or comments explaining non-obvious why (workarounds, constraints, subtle bugs).

---

Files in this repo

AGENTS.md                             ← Gemini CLI / any agent that reads AGENTS.md
.github/copilot-instructions.md       ← GitHub Copilot custom instructions
mcp/server.mjs                        ← MCP server for Claude Desktop, Cursor, Zed, Windsurf, VS Code

---

License

MIT