@graneth/mcp-server

@graneth/mcp-server

An account-free Model Context Protocol server that catches AI-hallucinated dependencies and hardcoded secrets in your staged changes before you commit — straight inside your AI coding agent.

It exposes a single, always-free tool over stdio:

pre_flight_check

Give it your staged files (path + content). It will:

• Detect hallucinated / slopsquatted packages — every package referenced in import statements and declared in manifests (package.json, requirements*.txt, Cargo.toml, go.mod, Gemfile, composer.json) is checked live against six registries: npm, PyPI, crates.io, RubyGems, the Go module proxy, and Packagist. A package that doesn't exist (404) is the hallmark of an AI-invented dependency an attacker may have pre-registered. → BLOCKED. A manifest that cannot be parsed is reported, never silently skipped, and an unreachable registry is reported as "could not verify" — never as clean. → REVIEW_REQUIRED.
• Flag brand-new packages (published < 30 days ago) — the prime slopsquatting attack window. → REVIEW_REQUIRED.
• Catch hardcoded secrets — known credential patterns (AWS, GitHub, Slack, OpenAI sk-…/sk-proj-…, Anthropic sk-ant-…, Stripe sk_live_…, PEM private keys) plus Shannon-entropy analysis with semantic variable-name context. → BLOCKED / REVIEW_REQUIRED. Vendor-documented sample keys and secrets in test/fixture files are downgraded to warnings — surfaced, but they won't block a commit.

No Graneth account, API key, or hosted backend is required. The checks run locally; the only network calls are to the public package registries (npm, PyPI, crates.io, RubyGems, proxy.golang.org, Packagist).

Install / run

npx -y @graneth/mcp-server

Configure your agent

The snippet is the same everywhere; only the file it goes in (or the CLI command) differs per client:

{
  "mcpServers": {
    "graneth": {
      "command": "npx",
      "args": ["-y", "@graneth/mcp-server"]
    }
  }
}

Client	Add it via	Config location
Claude Code (CLI)	claude mcp add --transport stdio graneth -- npx -y @graneth/mcp-server	Writes to ~/.claude.json (local scope, default). Add --scope project to write a shareable .mcp.json at the repo root instead.
Cursor	Paste the snippet above	~/.cursor/mcp.json (global) or <project-root>/.cursor/mcp.json (this project only)
Windsurf	Paste the snippet above	~/.codeium/windsurf/mcp_config.json (macOS/Linux) · %USERPROFILE%\.codeium\windsurf\mcp_config.json (Windows)
Claude Desktop	Paste the snippet above	macOS: ~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

Then ask your agent to run pre_flight_check before suggesting a commit. A BLOCKED verdict means: do not commit until the CRITICAL findings are fixed.

Verdicts

Verdict	Meaning
CLEAR	No issues — safe to commit.
REVIEW_REQUIRED	Warnings to confirm (new packages, lower-confidence).
BLOCKED	Critical issues (non-existent package / secret) — do not commit.

How it relates to Graneth

This server shares its detection core with the hosted Graneth scanner via the internal @graneth/core-checks module, so local pre-flight results match what the full PR scan would find.

License

MIT