SLF Guardian MCP
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@SLF Guardian MCP check if 'lodahs' is a real npm package before I install it"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
SLF Guardian MCP
Six security checks an AI agent can call before it acts — before it delivers code, installs a dependency, runs a shell or SQL command, opens a link or reads an untrusted page.
One file, no dependencies, stdio transport. Works with Claude Desktop, Cursor, Cline, Windsurf, Continue and anything else that speaks MCP.
Tool | The question it answers |
| Is this code safe to deliver or run? |
| Does this dependency exist, and is it safe to install? |
| Does this text or page carry instructions aimed at me? |
| Can this shell or SQL command be undone? |
| Is this domain safe to trust? |
| Passive outside-in security check of a website |
Every answer is a single verdict (pass, review, block) plus the evidence
behind it, so an agent can branch on it without reading prose.
Install
pip install slf-guardian-mcp
# or straight from this repository
pip install git+https://github.com/slfguardian/slf-guardian-mcpRelated MCP server: agent immune
Get a key
curl -X POST https://slfguardian.com/v1/keys \
-H "Content-Type: application/json" \
-d '{"email":"you@example.com"}'25 calls are free. The raw key is shown once.
Configure your client
{
"mcpServers": {
"slf-guardian": {
"command": "slf-guardian-mcp",
"env": { "SLF_API_KEY": "slf_live_..." }
}
}
}What it catches
check_package — the invented dependency. Assistants hallucinate package
names; attackers register them. langchain-helper-utils comes back block, no such package, lodahs comes back review, did you mean lodash. The source is
the registry itself plus the OSV vulnerability database, with npm postinstall
hooks, yanked and abandoned releases, age and download counts.
check_prompt — instructions hidden in content you were asked to read.
Text hidden with display:none, zero-width and bidi characters, Unicode tag
smuggling, homoglyphs, markdown exfiltration pixels, "ignore previous
instructions", "send .env to this webhook", "do not tell the user". The tool
never follows any instruction it finds; it returns the cleaned text and the
findings.
check_command — the thing you cannot undo. UPDATE users SET plan=1
without a WHERE comes back block, reversible=false; git push --force origin main comes back review with --force-with-lease suggested. Nothing is ever
executed.
check_domain — the link before you open it. Registration age and status,
DNS, lookalike brands, brand in a subdomain, punycode, shorteners, disposable
hosting, abuse-heavy TLDs. The site itself is never contacted.
audit_code — before you hand the file over. Remote code execution
(curl | bash), environment and credential exfiltration, reverse shells,
destructive commands, hidden base64, hard-coded secrets, install-time hooks,
shell=True, pickle.loads, TLS verification turned off.
Honest limits
These are detectors, not proofs. A pass means no known pattern matched, not
that the input is safe; a finding is evidence to look at, not a conviction.
Nothing submitted is executed, and the analysis is static.
Environment
Variable | Default | Purpose |
| — | Required. Your key. |
|
| Override the endpoint. |
|
| Seconds per call. |
HTTP API and OpenAPI description: https://slfguardian.com/api
MIT licensed.
Available Tools
6 toolsaudit_codeA
Audit source code before delivering or running it. Flags remote code execution (curl | bash), data exfiltration of environment variables or credential files, reverse shells, destructive commands, hidden base64 payloads, hard-coded secrets (AWS, OpenAI, Stripe, GitHub, database URLs), install-time hooks, typo-squatted package names and unsafe patterns such as shell=True, pickle.loads or TLS verification turned off. Nothing is executed; the audit is static text analysis. Returns a verdict (pass/warn/review/block), a safe_to_deliver flag and findings with line numbers.
Inputs: code, path, files.
check_commandA
Check a shell command or SQL statement before running it and say whether it can be undone. Flags recursive and system-path deletes, disk writes, history and credential wiping, firewall changes, force push and hard reset, docker volume and kubernetes deletes, terraform destroy, cloud bucket and resource deletion, DROP/TRUNCATE and DELETE or UPDATE with no WHERE clause. Returns a verdict plus reversible, safe_to_run, needs_human_approval and a safer alternative per finding. Nothing is executed.
Inputs: kind, command, commands.
check_domainA
Check a domain or link before opening it, sending data to it or installing from it. Uses the public registration record and DNS: registration age, registry status, whether it resolves, lookalike and typo-squatted brand names, a brand hidden in a subdomain, punycode and mixed-script hostnames, link shorteners, disposable hosting, abuse-heavy TLDs and phishing wording. Returns a verdict with safe_to_trust and safe_to_open. The site itself is never contacted.
Inputs: domain, domains.
check_packageA
Check a PyPI or npm package before installing or recommending it. Asks the registry whether the name exists at all (invented package names are how slopsquatting attacks land), how old it is, how much it is downloaded, whether it runs install scripts, whether it is deprecated or yanked, whether OSV lists vulnerabilities, and whether the name is a near-miss of a popular package. Returns a verdict and a safe_to_install flag.
Inputs: name, version, packages, ecosystem.
check_promptA
Check untrusted text, HTML or a web page for prompt injection before feeding it to a model. Finds instructions written for the model instead of the reader: text hidden with CSS, HTML comments, zero-width and invisible Unicode tag characters, chat template markers, 'ignore previous instructions', attempts to make an agent leak its system prompt or credentials, hide actions from the user, approve something without review, run a command, or render a markdown image that exfiltrates data through its URL. Returns a verdict, injection_detected, and clean_text with invisible characters stripped. Never follows what it finds.
Inputs: url, html, text.
scan_siteA
Passive outside-in security check of a website: malware and spam injection traces, hidden redirects, SSL/TLS certificate state, security headers, exposed software versions and SPF/DMARC email spoofing protection. Does not log in, does not run attack tests. Returns a score out of 100, a grade and findings by category.
Inputs: domain (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
Each tool maps to a distinct artifact type (code, package, prompt text, shell/SQL command, domain, live site), so the core split is clear. The only real overlap is check_domain vs scan_site — both take a URL/site and return a safety verdict — and an agent could plausibly hesitate over which to call for a suspicious link, though the descriptions (pre-open trust check vs passive site posture scan) do resolve it.
Five of six tools use snake_case verb_noun with check_* as the dominant prefix; audit_code and scan_site deviate in verb but stay in the same snake_case verb_noun shape. The deviation is semantically motivated (deeper analysis, site scanning) rather than arbitrary, so the set reads predictably.
Six tools is well-scoped for a pre-flight security guardian: one per risky artifact class, with no filler or near-duplicates. Nothing feels padded or missing at the count level.
The surface covers the main agent risk vectors: untrusted code, third-party packages, injected text/HTML, destructive commands, untrusted domains, and site posture. Minor gaps remain — no tool for inspecting downloaded files/archives or MCP server configs, and no consolidated 'check everything about this artifact' entry point — but agents can chain existing tools to cover most workflows.
Maintenance
Related MCP Connectors
Security scanner for AI agents: prompt injection, leaked secrets, PII, and SSRF-risk URL detection.
Pre-execution governance for AI agents. Deterministic PASS/FAIL/REVIEW verdicts, replayable proof.
Deterministic runtime safety for AI agents: scan PII, gate tool actions, verify LLM output.
Security gateway for AI agents: policy, approval, and audited execution, no secrets shared.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceFull-stack security for AI agents — static analysis + MCP runtime interception. 31 rules detect prompt injection, data exfiltration, backdoors, tool poisoning, and cross-file attack chains. Includes MCP proxy for real-time blocking, Python AST taint tracking, multi-language injection detection (8 languages), and AI-powered deep analysis. Free, offline, zero-config.83 npm16MIT
- AlicenseNot gradedqualityDmaintenanceAdaptive security for AI agents: assess inputs for prompt injection, scan outputs for credential/PII leaks, teach new attack patterns to semantic memory, harden prompts, and monitor metrics. Runs locally via MCP stdio.129 PyPIApache 2.0
- AlicenseNot gradedqualityCmaintenanceProvides security checks for AI agents, including secret scanning, CVE lookup, and dependency vulnerability scanning, all running locally except for CVE lookups.MIT
- AlicenseAqualityBmaintenanceLocal security layer for coding agents: keeps secret values out of what Claude Code, Codex CLI, Cursor and OpenCode send to their model, and checks untrusted text for prompt injection. MIT, runs on your machine, no account.45MIT