A quick security X-ray for AI agent plugins, to run before you plug one in.

An MCP server (MCP = Model Context Protocol, the standard way to give an AI assistant new tools) is code you download and let an AI agent run. mcpaudit reads that code before you trust it and points out the dangerous bits — the quick safety check that doesn't really exist for these plugins yet.

npx allenwu-blip/mcpaudit ./path-to-an-mcp-server

No install, no setup, no API key, no internet needed. It reads the plugin's source code and its settings file and flags risky patterns, ranked by how bad they are, each with a concrete fix. It never runs the code it is checking — it only reads it.

Why

These plugins run with real power inside the AI agent's loop — they can get a shell, your files, and the network on your machine, and whatever a plugin's tools output flows straight back into the AI's context where it can steer what the AI does next. Published MCP servers get inconsistent security review before people wire them in — the 2026 surveys from dev.to and The Register give differing pictures of how widespread the risk is — but the failure modes are concrete: command injection, environment/credential leakage into LLM-visible context, and over-broad filesystem/tool scope.

mcpaudit exists so you can run a concrete check for those patterns yourself, in seconds, offline, before letting someone else's plugin run inside your agent.

Install / run

It's zero-install via npx. A local path is scanned fully offline:

# scan a server you cloned / vendored
npx allenwu-blip/mcpaudit ./vendor/some-mcp-server

# machine-readable output for CI / tooling
npx allenwu-blip/mcpaudit ./server --json

# SARIF v2.1.0 (a standard scan-results format GitHub understands) —
# upload it so findings show in GitHub's Code scanning tab
npx allenwu-blip/mcpaudit ./server --sarif > mcpaudit.sarif

# stricter gate: any high or critical fails the command
npx allenwu-blip/mcpaudit ./server --fail-on high

# continuous monitoring: accept current state, then gate only on NEW
# regressions (offline, no accounts) — see "Continuous monitoring" below
npx allenwu-blip/mcpaudit ./server --baseline-write .mcpaudit-baseline.json
npx allenwu-blip/mcpaudit ./server --baseline .mcpaudit-baseline.json

Scanning by bare package name (npx mcpaudit some-mcp-pkg) needs a registry/tarball fetch wired up; the published build asks you to pass a path instead (it does not silently do nothing and does not hit the network). The path scan is the fully-functional path today.

Exit codes (so you can wire it into CI)

CI ("continuous integration" — the automated checks that run on every code push) reads these exit codes:

Fails open on purpose: if mcpaudit itself breaks (a bug, a folder it can't read), it prints a loud error and exits 0. A security checker that is itself broken should not block every build in your project. If you want it to be a hard stop, make it a required check with --fail-on set, so a missing or zero result is visible rather than silently passing.

What it detects

This is the detailed reference for developers. The rules are fixed and give the same answer every time (no AI, no guessing). Each finding has a stable id, a severity (how serious), the exact file:line:col location, a plain explanation of why it fired, and how to fix it.

The rules are intentionally conservative — they aim to avoid the obvious false-positive patterns (literal exec/eval, env used only for auth, scoped relative directories, fixed-origin URLs, path.join-contained file access, safe-schema YAML, placeholder secrets, normal build hooks, and scary tokens that are only in comments). The bundled borderline fixture is a legit MCP server full of code that looks dangerous and must produce zero findings; it is part of CI.

Output formats

Continuous monitoring (baseline diff — free, offline, no accounts)

A one-shot scan tells you today's state. A team usually wants "did anything get worse since we last reviewed this server?" That is a diff against a committed baseline — pure, deterministic, offline, no sign-up:

# 1. accept the current state into a baseline and commit it
npx allenwu-blip/mcpaudit ./server --baseline-write .mcpaudit-baseline.json
git add .mcpaudit-baseline.json && git commit -m "mcpaudit baseline"

# 2. in CI: re-scan and gate ONLY on NEW findings (regressions). An
#    already-triaged finding no longer re-breaks every build; a freshly
#    introduced one does.
npx allenwu-blip/mcpaudit ./server --baseline .mcpaudit-baseline.json --fail-on high

The baseline file is intentionally timestamp/host/user-free so re-writing an unchanged repo is byte-identical (clean, reviewable PR diffs); when a finding appeared is git's job, not the file's.

Note (honest scope): the hosted/continuous-monitoring product — a service that watches a server over time, alerts on a new critical, or shows a fleet view — is not in this repo. This OSS CLI ships only the baseline-diff mechanic and emits the machine record a hosted tier would consume (--monitor-json). There is no network call, no upload, no account, no billing anywhere in this codebase, by design.

GitHub Action

A thin wrapper around the same scan. Copy examples/mcpaudit.yml into .github/workflows/:

- uses: allenwu-blip/mcpaudit@v0
  with:
    path: "."
    fail-on: "high"
    sarif: "true"          # optional: write mcpaudit.sarif
    # baseline: ".mcpaudit-baseline.json"  # optional: gate on NEW only

It posts a GitHub annotation per finding and sets outputs (total, critical, high, medium, low, gate, plus new/fixed with a baseline and sarif-file with sarif: true). With sarif: true it writes a SARIF v2.1.0 file you upload via github/codeql-action/upload-sarif@v3 (see examples/mcpaudit.yml for the security-events: write permission and upload step). No token or secret is needed for the scan itself; the Action does no network I/O and never runs the scanned code. It fails open on internal error — make the job a required check for hard enforcement.

Limitations (read this)

mcpaudit reads code and matches known-dangerous patterns. It does not run the code in a locked box (a "sandbox") and it does not trace exactly how a value flows from input to a dangerous spot ("taint analysis"). Be clear-eyed about what that means:

• It will miss things (false negatives). Obfuscated code, vulnerability reached through indirection/aliasing, dynamic require, behaviour that only manifests at runtime, or a malicious dependency several layers deep are largely out of scope. A clean result is not a security guarantee or an audit — review the server's tools and scope yourself.
• It does not do taint tracking. It cannot prove a sink is reachable from tool input; it flags the dangerous shape and tells you to verify reachability. Conversely it deliberately under-reports to stay quiet: e.g. MCP010 does not flag path.join("./dir", x) even though a .. in x can still escape — full path-containment analysis is beyond a lexical scanner, so that is a known, accepted false negative, not a guarantee the call is safe. Treat every path.*/merge/fetch on tool input as worth a human look regardless of whether a rule fired.
• Provenance-gated, conservative by design (favor a false negative over cry-wolf). A few rules deliberately stay quiet on idiomatic-safe shapes:
  • MCP005 fires only on a bare global eval( / new Function(, and on the vm builtin's runInThisContext/runInNewContext/runInContext/ compileFunction. A method call that merely shares those names — mathExpr.compile(x), parser.eval(x), engine.compile(tmpl), an ORM .run() — does not fire (it is not the global / vm). The vm.* sink is suppressed when vm is provably a userland binding (e.g. const vm = makeSandboxShim()). Bare eval(userInput) still fires by design — that is the sink.
  • MCP010 also treats hoisted containment as safe: when the fs.* path is a bare variable whose nearest prior assignment is built from path.join/resolve/normalize/basename(...) (e.g. const safe = path.resolve(BASE, path.basename(name)); fs.readFileSync(safe)), it does not fire. This look-back is the nearest declaration/ assignment of that identifier and is intentionally single-hop — a containment value passed through additional indirection (further aliasing, a helper return) is a known, accepted false negative (same stance as the path.join note above), not a safety guarantee.
• It will sometimes be wrong (false positives). The rules use a lexical model (comments and static string text are excluded; template interpolation is treated as code). Unusual code can still trip a rule. Please report misfires — that is how it improves.
• No accuracy/benchmark numbers are claimed. There is no published labeled corpus behind this tool, so it ships with no precision/recall or detection-rate figures, and it makes no claim about how it compares to any external survey of the MCP ecosystem.
• The dependency layer is static and offline — and makes no CVE claim. MCP014 flags non-registry dependency sources and, as a low advisory only, names that are one edit from a popular package. It contacts no registry, bundles no vulnerability database, and asserts nothing about whether a given package/version is malicious or has a known CVE. Pair it with a real SCA/advisory tool (npm audit, OSV, Dependabot). Registry/tarball fetching for mcpaudit <name> remains an unimplemented, honest interface — the published build asks you to pass a path.
• Scope is the documented MCP/Node patterns. JS/TS source + JSON manifests. Servers written in other languages, or that hide configuration outside the manifest, are not fully covered. Minified/bundled, binary/non-UTF8, and symlinked files are deliberately skipped (and surfaced as diagnostics) — audit the original source, not build artifacts.
• It is one layer. Use it alongside dependency scanning, least-privilege configuration, and human review — not instead of them.

How it works

• src/rules.js — the pure, deterministic analyzer (no I/O, no network, no code execution): 14 source/manifest rules. Fully unit-tested.
• src/analyze.js — the only filesystem touch: walks the tree, never follows symlinks (cannot be steered out of the target), skips minified/binary/non-UTF8 blobs and pathological depth, never throws (errors are collected and returned), and never runs the project.
• src/format.js — pure presentation: human, --json, and SARIF v2.1.0.
• src/baseline.js — pure baseline build + diff for continuous monitoring; contains the documented // PAID TIER seam (no network/account code).
• bin/mcpaudit.js / src/action.js — thin CLI / Action glue.

Finding ids are a deterministic hash of rule|file|line|col|message, so two distinct findings at the same location stay distinct and the id is stable across runs and machines — CI baselines, SARIF de-dup, and suppression lists are reproducible.

Development

npm ci
npm test          # vitest — no network, no API key required
node bin/mcpaudit.js test/fixtures/vulnerable-server   # try it
node bin/mcpaudit.js test/fixtures/vulnerable-server --sarif | head

Tests run against three fixture MCP servers — a clean one (zero findings), a vulnerable one (every one of the 14 rules fires with the right severity and location), and a borderline one (legit code that looks scary across all 14 rules and must stay at zero findings) — plus dedicated suites for SARIF schema correctness, baseline-diff behaviour, and adversarial hardening (symlink escape, minified/binary/non-UTF8, deep trees, never-throws).

Feedback

False positives and missed vulns are the most valuable input. See FEEDBACK.md: add the mcpaudit-feedback label to an issue, or use the issue template. Reports are captured verbatim — read exactly as written, never paraphrased.

License

MIT — see LICENSE.