Security scanner for AI-generated code — the check a coding agent runs on its own code or diff before it commits. Deterministic, free, no LLM. Live: https://code-guard-api.vercel.app
A coding agent writes a patch and commits it in the same breath. Nothing cheap sits in that loop and reads the diff first. Code Guard is that first-line scanner: a deterministic rule engine for the high-frequency classes. It is not a replacement for a real audit.
Detects: command / code / SQL injection · SSRF (heuristic) · hardcoded secrets & API keys · weak crypto
(MD5/SHA1, ECB, DES, insecure RNG) · unsafe deserialization (pickle/yaml/marshal/node-serialize) · disabled TLS
verification · XSS / server-side template injection · misconfig (Flask debug, JWT alg=none).
31 code rules across 12 categories, each with a stable ID, plus hardcoded-secret detection covering 12
credential patterns and 3 personal-data patterns (email, US SSN, card number), all under hardcoded-* ids.
GET /api/rules returns 32 entries because it lists the secret and PII rules as one grouped hardcoded-*
entry. Every scan response carries the rules_version that produced it.
Languages: JS/TS and Python. That is the whole list. 11 rules are JS/TS, 14 are Python, 6 are
language-agnostic. There is no Go, Ruby, PHP, Java or Rust ruleset. A language I don't recognise is
ignored. I sniff the source instead and land on one of three answers, js, py or unknown. Which one
you get depends on what the file happens to contain. import java.util.List; reads as Python. A Go
file with a const line reads as JavaScript. A Go file with nothing but func and import "fmt" lands
on unknown, and then all 31 rules run against every line, which catches more and flags more things that
are not bugs. Outside JS/TS and Python, count on the 6 language-agnostic rules and the secret scan.
Anything else you get is luck.
The 31 code rules are regexes matched one line at a time. No parser, no data flow, no taint tracking, so
it isn't static analysis in the sense a SAST vendor means it. That cuts both ways.
db.query("SELECT ... " + id) is caught on one line and missed the moment you split it across two. The
word DES on a line trips weak-cipher, comment or variable name included (DES_KEY does not, the rule is
whole-word). The hardcoded-* patterns are the exception: those run over the whole source, so a PEM
private-key block spanning several lines is still caught. Treat a clean result as "none of my 31 rules
matched", not as "this code is safe".
curl -s -X POST https://code-guard-api.vercel.app/api/scan-code -H 'Content-Type: application/json' \
-d '{"language":"python","code":"os.system(\"echo \"+x)\nobj=pickle.loads(d)"}'
# { "verdict":"block", "counts":{"critical":2,...}, "findings":[{ "id":"py-os-system","line":1,... }] }Deterministic (same input → same output). Reuses the Agent Firewall secret-scanner. MCP server:
code-guard-mcp. Part of the agent-guardrail suite:
Package Guard · Agent Firewall · Payment Guard · Email Guard · Code Guard.