flagrix MCP server

flagrix

Scan GitHub repositories and profiles for malware before you clone — from the terminal, CI, or an AI agent. The same commit-pinned verdict as the Flagrix browser extension, made callable.

npx flagrix scan https://github.com/some-org/coding-assignment

  some-org/coding-assignment @ 3f9c2a1
  HIGH RISK — Do not clone  security score 12/100
  3 files scanned · 10 dependencies · 2 issues

  CRITICAL Data exfiltration patterns detected: Keylogger Pattern
    assignment.js:14
      14  document.addEventListener("keydown", (e) => send(e.key))

Built after real fake-recruiter campaigns ("coding assignment" repos that steal wallets, SSH keys, and browser sessions) started targeting developers.

Commands

flagrix scan <url | owner/repo>   # scan a repository (--ref <branch|sha>)
flagrix scan-user <username>      # score a GitHub profile for scam signals
flagrix mcp                       # MCP server (stdio) for AI agents

Exit codes

code	meaning
0	low risk
1	scan failed
2	medium risk — review before proceeding
3	high risk — do not clone
4	scan incomplete — not enough of the repository was read to give a verdict

--json (automatic when stdout is piped) emits the full result. The verdict is pinned to the scanned commit (commitSha in the JSON): every file is read at that SHA, so a push mid-scan or after the verdict can't silently invalidate it.

Incomplete scans (exit 4)

A repository scan is one tree request plus up to ~200 file reads, and GitHub allows 60 requests an hour without a token. When the budget runs out part way through, the rest of the repository is never looked at — so a "low risk, no findings" result over 30 of 180 files is the absence of looking, not a clean bill of health.

The result says so (scanSummary.complete: false, with scanSummary.incompleteReason of "rate-limit" or "fetch-errors"), the human output withholds the verdict and the score, and the exit code is 4 instead of 0. In CI the exit code is the decision and nobody reads stdout, so a gate on flagrix scan stops here as it does on any non-zero code. Findings are the other way round: evidence found is evidence however little else was read, so a medium or high verdict keeps its own code and the coverage note reads "there may be more".

In --json output, read verdict and coverage rather than riskLevel:

{
  "verdict": "incomplete",
  "coverage": {
    "complete": false,
    "unreliableClear": true,
    "filesRead": 30,
    "filesUnread": 150,
    "reason": "rate-limit"
  },
  "riskLevel": "low",
  "safeToClone": false,
  ...
}

• verdict is "low", "medium", "high", or "incomplete" when a low result rests on files that were never read. It matches the exit code (0 / 2 / 3 / 4).
• coverage.unreliableClear is true exactly when verdict is "incomplete"; coverage.complete is false on any partial scan, including a medium or high one. reason is present only when coverage is incomplete.
• riskLevel, riskScore and the rest of the engine's result are unchanged, so existing consumers keep working — but riskLevel is the engine's reading of the files it did see, so on an incomplete scan it can still say "low".
• safeToClone is kept for compatibility and is now false whenever verdict is "incomplete" (in 0.4.0 and earlier it could be true there). Gate on verdict or the exit code instead.

The MCP tool's JSON block carries the same fields.

Set a token to make this rare: GITHUB_TOKEN raises the budget to 5,000/h.

AI agents

claude mcp add flagrix -- npx -y flagrix mcp

Tools: scan_github_repo, scan_github_user. A Claude Code hook that gates every git clone on a scan ships in hooks/ — see docs/agent-gating.md.

Tokens & rate limits

Unauthenticated scans use GitHub's 60 req/h budget (a scan issues one request per scanned file, up to ~200). Set GITHUB_TOKEN (or FLAGRIX_GITHUB_TOKEN, or --token) to raise it to 5,000/h and to scan private repositories. A scan that runs out of budget reports partial coverage and exits 4 — see Incomplete scans.

Privacy

Fully local. No telemetry, no accounts, no Flagrix backend — the only network calls go to the GitHub/npm APIs and the public detection-rules repository (signature refresh, cached 6 h, with a bundled offline snapshot).

Dependencies are also checked against the OSV malicious-package feed (osv-malicious.json from the same repository: every npm/PyPI package the OpenSSF malicious-packages project has confirmed malicious). The feed is downloaded and matched locally, so dependency names never leave your machine. It is cached 6 h and revalidated with its ETag. Set FLAGRIX_OSV_FEED=0 to skip it (the curated list in the signatures still applies).

How it works

Scanning logic lives in @flagrix/scanner-core (MIT), signatures in flagrix-detection-rules (MIT) — the same engine and rules the browser extension uses. Verdicts are risk assessments, not definitive fraud determinations; always verify through official channels.

AI Disclosure

This project leverages Claude AI for boilerplate generation, test-suite expansion, and optimization. All AI-generated code is strictly reviewed, refactored, and verified by human maintainers before merging.

License

MIT — see LICENSE.