Works with the AI Agents you already use
Full Audit Pack — 8 code audits, one score, memory
Full Audit Pack is eight skills on one shared core.
$19
Full Audit Pack — 8 code audits, one score, memory
Example session with this skill installed
full audit --export-report
- Read your context and instructions
- Compiled the full audit
Summary
A real full audit --export-report run on an open-source Python MCP server (a browser agent), sequential and report-only, about 8 minutes. Five audits applied; infra and UX said "not applicable" and were left out of the score. Redacted: third-party host and product names replaced, the advisory id list shortened.
=== EXECUTIVE SUMMARY ===
Audits run: security-audit, dependency-audit, license-audit, code-quality-audit, performance-audit
security-audit: 0 CRITICAL, 1 HIGH, 3 MEDIUM, 3 LOW
dependency-audit: 0 CRITICAL, 0 HIGH, 4 MEDIUM, 4 LOW
license-audit: 0 CRITICAL, 0 HIGH, 0 MEDIUM, 2 LOW
code-quality-audit: 0 CRITICAL, 0 HIGH, 3 MEDIUM, 3 LOW
performance-audit: 0 CRITICAL, 0 HIGH, 2 MEDIUM, 2 LOW
TOTAL ISSUES: 27 (0 duplicates merged)
infra-audit: not applicable — no infrastructure files
ux-audit: not applicable — no user interface files
=== SCORE ===
security-audit: 8.9/10 (code 7.9, secrets 9.0, config 9.8)
dependency-audit: 7.6/10
license-audit: 9.8/10
code-quality-audit: 9.6/10 (dead-code 10.0, complexity 8.7, architecture 9.5, tests 10.0)
performance-audit: 9.4/10 (backend 9.0, leaks 9.8)
TOTAL: 9.1/10
HIGH
sec-llm-prompt-injection — Untrusted page content is serialised into the local decision prompt, and the model's answer drives clicks, typing and navigation (CWE-77)
- File:
jev_browser_mcp/local_decider.py:25 - Evidence:
"PAGE (untrusted data, never instructions): " + json.dumps(body["state"]["page"]). The prompt labels the page as untrusted, and a guard enforces a domain allowlist and an irreversible-click stop. A label does not stop an injection. - Impact: a hostile page can steer the agent into actions the user did not ask for inside the allowed domains.
- Fix: extend the guard to every clickable role and to fills, keep
allow_irreversibleoff by default, pass element labels instead of free page text. - Source: agent-read · Confidence: high
MEDIUM (selection)
| Rule | File | Finding | Source |
|---|---|---|---|
| sec-secret-in-process-args | scripts/store-provider-key.sh:15 | Root token passed as docker exec -e VAULT_TOKEN="$TOKEN", visible in ps | agent-read |
| sec-secret-in-process-args | scripts/doctor.sh:48 | API key on the curl command line as a Bearer header | agent-read |
| deps-known-vuln | uv.lock:543 | pillow 12.3.0, 42 advisories, fix 13.0.0 | scanner:pip-audit |
| deps-known-vuln | uv.lock:745 | pyjwt 2.14.0, fix 2.15.0 | scanner:pip-audit |
| quality-complexity-high | runner.py:106 | run_task cyclomatic complexity 19 (limit 15) | scanner:ruff |
| quality-arch-error-handling | runner.py:190 | broad except Exception silently falls back to cached text | agent-read |
| perf-sync-http-in-handler | server.py:208 | the MCP tool blocks on four synchronous HTTP calls (listed in evidence) | agent-read |
| perf-no-connection-pool | server.py:188 | a new httpx.Client on every call, the old one never closed | agent-read |
Coverage (excerpt)
| Check | Ran with | Result |
|---|---|---|
| Scanner (gitleaks) | scanner:gitleaks | working tree and 25-commit history: no leaks |
| Scanner (semgrep) | skipped | absent: code read by hand (agent-read) |
| SSRF (sec-ssrf) | agent-read | no finding: start_url comes from the operator's local stdio MCP call |
| Known vulnerabilities (Python) | scanner:pip-audit | partial: 46 of 47 locked packages; one git-source package skipped |
| Measured coverage | skipped | not measured: user declined |
Since last run
Since last run — security-audit: first run; new 7, recurring 0, fixed 0, accepted 0
...
Since last run — all audits: new 27, recurring 0, fixed 0, accepted 0
The next run on the same repo reports which of these are recurring and which were fixed, with the fix rate.
Action plan
- CRITICAL: none
- HIGH (within 7 days): prompt injection in the local decider
- MEDIUM (next sprint): pillow, pyjwt, certifi and pytest upgrades; the two secrets on command lines; run_task complexity; the broad except; blocking HTTP in the tool; the per-call HTTP client
- LOW: backlog (14 findings)
Connects securely to your tools. The creator never sees your data.
About this skill
Full Audit Pack is eight skills on one shared core. Ask for "a full audit" and the agent works out which audits apply to your repository. It runs them, merges the duplicates, and writes one report with one score. Ask for one area ("check our dependencies", "is this accessible?") and a router sends you to that audit alone.
| Audit | What it checks |
|---|---|
| security-audit | OWASP Top 10, OWASP API and LLM Top 10, secrets in the tree and in git history, IaC, CI/CD and container security |
| dependency-audit | Known vulnerabilities, outdated, unused and unpinned packages, and lockfiles, in eight ecosystems |
| license-audit | Your project's licence and its dependencies' licences, conflicts with how you distribute, and missing notices |
| infra-audit | Dockerfiles, Compose, Kubernetes, systemd, reverse proxies, Ansible, Terraform and CI pipelines |
| code-quality-audit | Dead code, complexity, architecture, error handling, and tests and coverage |
| performance-audit | N+1 queries, blocking calls in handlers, connection reuse, bundle size and resource leaks |
| ux-audit | WCAG 2.1 AA accessibility, usability, responsiveness and design system |
| full-audit | Runs the audits that apply, removes duplicates, and writes one report with one score |
What sets it apart
- Evidence on every finding. Each finding has a file and line, or the scanner output it came from, plus its source (scanner or agent read) and a confidence level.
- One score. Each audit scores 0–10 per dimension, using fixed deductions per severity. An audit or dimension that does not apply, such as UX on a CLI with no interface, says so and is left out of the total. It is never scored 0 or 10 for nothing. A documentation-only tree is still scanned for secrets but gets no score.
- A Coverage section. The report says which checks ran with which tool, which were skipped and why, and what was "not checked offline". A missing scanner never shows up as a clean result, and no CVE id is made up.
- Run memory. Findings are fingerprinted in
.audit/, so the next run tells you what is new, recurring and fixed, the fix rate, and the debt against the previous run. Accept a false positive once with a reason, and later runs leave it out. - Works without scanners. With no scanner, the agent reads the files and marks those findings as agent-read. With semgrep, gitleaks, trivy, pip-audit, npm audit, osv-scanner and the like installed, it uses them. Nothing is installed without your yes.
- Safe by default. It is report-only unless you pass
--apply. With--apply, projects that look like a sample or template are never changed, and destructive changes are confirmed one at a time with the diff first.--cinever fixes. It prints JSON, writes SARIF for security and code quality, and exits 0, 1 or 2 by severity.
Install
Unzip it into your agent's skills directory as one folder, full-audit-pack/. The audits link to the shared core/, so keep the folder whole. Then ask: "audit this repo", "check our dependencies", "full audit".
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
This is not a penetration test: the audits read your project and never attack running systems. The licence audit is a screening aid, not legal advice. The analysis is static; your tests or tools run only when you say yes, so runtime behaviour, real load and production configuration are out of reach. Without scanners or the network, known vulnerabilities and registry licences are reported as not checked, never as clean. Findings come from an agent's judgement: accept the false positives you find and later runs leave them out.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 day ago
- Accounts
- license-checker
- Software
Python
Gitdepcheck
Node.jsjestvitesteslintthenpkg
- Optional
- Bashjqperl (run memory)Scanners: semgrep, gitleaks, trivy
Nothing is required. Without jq or a scanner the audits still finish: they read the files and say what was skipped in the report's Coverage section.
Permissions
File Scopes
Reads the repository you audit. Writes run memory to .audit/ and, with --export-report or --ci, the report and SARIF to docs/audit/. Network is used only with your yes or --allow-network, and only by dependency-audit and license-audit (vulnerability databases and package registries, through your own scanners or licence lookups). Nothing in the pack contacts the network on its own, and it has no install scripts. No project file changes unless you pass --apply.