AgentGuardGuardrails and allowlists before agents ship.
Map controls to NIST RMF themes and lock tool allowlists — decision-support, not a security audit certificate.
No credit card required · Cancel anytime
Key Takeaways
- NIST RMF-mapped guardrail checklist.
- Tool allowlists for OWASP agent risks.
- Honesty: decision-support — not a pentest or conformity certificate.
Guardrails first. Allowlists next.
Decision-support themes — not a conformity certificate.
AgentGuard — definition
AgentGuard is a pre-deployment guardrail review: it scores an agent system prompt and declared tool list for secret leakage, prompt injection, and unsafe tool use (rule-based checks plus optional model-assisted notes) and returns a cited report. It does not intercept live traffic like Lakera Guard or NVIDIA NeMo Guardrails.
What this is NOT
- Not a law firm, GRC platform, or ISO certification body.
- Not a WAF, a runtime sidecar, or a live traffic interceptor.
- Not a SIEM, pentest suite, or conformity certificate.
Controller / Processor — When packs mention privacy roles, we use GDPR Art.4 labels: Controller and Processor (decision-support only).
Open tools. Open risk.
No allowlist
Agents call tools they should not.
Unmapped controls
NIST themes never reach the runtime.
Late reviews
Security shows up after ship.
Overclaim tools
Promises of certified safety erode trust.
Describe. Review. Lock.
Decision-support guardrails — not a pentest certificate.
NIST-mapped checklist
Themes you can discuss.
Tool allowlists
Reduce OWASP exposure.
Ship-ready review
Built for pre-prod gates.
Honest demo vs live
Demo is labelled; live needs a configured key where applicable.
Everything you need for a first-pass guardrail review
Tool-risk checklist
A deterministic checklist with a versioned ruleset, so every run scores the same declared tool list the same way.
Tool-call risk rating
Per-tool severity and risk appetite modes (Strict / Balanced / Permissive) with blocklist suggestions.
EU AI Act mapping
Input sanitisation and output filter ideas, plus side-effect gates on the tools that change state.
Blocklist suggestions
EU AI Act theme pointers and an exportable report you can attach to a pre-prod review.
From policy to enforced guardrails in three steps
AgentGuard follows the same workflow every run — deterministic, versioned, and exportable.
Paste your agent setup
Provide the agent system prompt, the tool list, and your risk appetite — no integration required.
- System prompt and tool declarations as input
- Risk appetite: Strict, Balanced, or Permissive
- Works with any framework that produces a prompt
Rate every tool call
Each declared tool is scored for risk and matched against your appetite, with input/output filter proposals.
- Per-tool severity with file-and-line citations
- Input sanitization and output filter proposals
- Side-effect gates on irreversible actions
Apply and export the ruleset
Turn findings into blocklists and a versioned guardrail report you can gate pre-prod builds on.
- Blocklist suggestions per tool
- Exportable report with rulesetVersion
- Same ruleset every run — auditable
See how teams use AgentGuard
Different agents, different risks — the same deterministic checklist.
Email, SQL and HTTP tools that can take irreversible actions
AgentGuard rates each tool, proposes I/O filters, and suggests blocklists before the agent ever runs in production.
Try it on this page
or generate below.
Run history & export
What the studio actually reviews
Capability cards, not customer quotes. You still own the allowlist before production.
“A versioned ruleset scores the declared tools the same way on every run.”
“Draft filters and side-effect gates for tools that send mail, run SQL, or call HTTP.”
“Highest-risk tools can be tagged to Act themes. That is decision-support, not a conformity certificate.”
Honest boundaries
AgentGuard drafts a guardrail review. No pentest or conformity certificate.
Simple plans that scale
All paid plans include
- Guardrail studio
- NIST-mapped checklist
- EU AI Act theme mapping (decision-support)
- Email support on paid plans — no hidden add-ons for core studio use
Try the full three-step workflow before committing.
- 3 AI runs/day · 15/mo (enforced fair use)
- Studio demo (no extra LLM fee)
- No audit log
- Watermarked export only
- No BYOK
For agents heading to production — clean exports and an audit trail.
- AI drafts 40/day · 300/mo
- No separate ChatGPT subscription required
- Priority support · 2 months free with yearly
For teams that need higher caps and their own LLM key.
- Team seats · API access (roadmap)
- Optional BYOK (your OpenAI key)
Feature comparison
What each plan includes — quotas reflect the enforced fair-use limits.
| Dimension | Free | Pro | Enterprise |
|---|---|---|---|
| Usage | |||
| AI runs | 3/day · 15/mo | 40/day · 300/mo | 200/day · 3,000/mo |
| Rule-only checks (no AI) | 5/day · 50/mo | 200/day · 3,000/mo | 1,000/day · 30,000/mo |
| Report export | Watermarked | Clean export | Clean export |
| Audit log | Not included | ✓ Included | ✓ Included |
| Guardrails | |||
| Tool-risk rating | ✓ Included | ✓ Included | ✓ Included |
| I/O filter proposals | ✓ Included | ✓ Included | ✓ Included |
| EU AI Act mapping | ✓ Included | ✓ Included | ✓ Included |
| Blocklist suggestions | ✓ Included | ✓ Included | ✓ Included |
| BYOK (your LLM key) | Not included | Not included | ✓ |
| SSO | Not included | Not included | Roadmap |
| Support & compliance | |||
| Support | Community | Priority | Custom SLA |
Already paid? Unlock Pro with your license
Licenses are issued manually after your payment is confirmed: we generate one per paid order and send it to the purchase email, typically within a few hours. Enter it below to activate Pro on this device — no account needed. Nothing yet? Write to lixingliangsy@163.com and reference your order email.
Your questions, answered
What does AgentGuard check?
Input/output guardrails and tool allowlists with NIST RMF / OWASP framing — decision-support, not a pentest certificate.
Is this legal advice or a certification?
No. Outputs are decision-support drafts. Have counsel and risk owners review before production. We do not claim ISO, CMP, or regulator certification.
Can I cancel anytime?
Yes. Self-serve plans cancel anytime; access continues until period end.
Do I need a credit card to start?
No. Start with email signup or Demo mode in the studio, then upgrade when ready.
Do I need my own ChatGPT / OpenAI subscription?
No for Free/Pro. AI runs are included in your plan (fair use) via our platform key. Enterprise can optionally bring your own OpenAI key (BYOK) — configure it at /settings (key stays server-side only).
What is Fair Use / what if I hit the AI limit?
Pro includes 40 AI drafts per day and 300 per month. If you hit the fair-use limit, Studio returns 429 QUOTA_EXCEEDED (no silent mock) until reset — or upgrade / use Enterprise BYOK.
Is checkout secure?
Payments are processed by Waffo Pancake (merchant of record).
What happens after I pay?
You receive access confirmation; fulfillment is tracked via webhook + order logs.
Quick answers
- Tool-risk checklist
- Tool-call risk rating
- EU AI Act mapping
- Blocklist suggestions
- Pricing starts at $0 (Free).
More questions
What risks does AgentGuard catch?
Secret leakage in prompts, prompt-injection patterns, and unsafe tool or function calls, mapped to the OWASP Top 10 for LLM Applications.
Does it run live AI?
Pro uses the platform LLM key under fair use; Free and demo mode return a labelled Rule-based report. Missing keys or upstream failures return 503/502 — never a silent mock.
Is it a WAF or runtime guard?
No. It is a pre-deployment review of the agent prompt and tool list. It is not Lakera Guard, NVIDIA NeMo Guardrails, or a network firewall.
Does it map to the EU AI Act?
Highest-risk tool classes can be tagged to relevant AI Act themes for documentation. That mapping is decision-support, not a conformity assessment.
Who is AgentGuard for?
Teams shipping agents who need a documented guardrail review before launch.
How it compares
| Dimension | Manual | AgentGuard |
|---|---|---|
| Speed | Hours to days | Minutes per run |
| Consistency | Varies by person | Same ruleset every run |
| Output | Free-form | Structured, exportable result |
| Best for | Final sign-off | First-pass decision-support |
When not to use: Use qualified human review instead — it is a drafting aid. Policies must be reviewed by your counsel and risk owners.
- Pre-prod guardrail review — Score the prompt and tool list before the agent ships.
- EU AI Act theme tags — Tag the highest-risk tools to Act themes. Not a conformity certificate.
- Wave 1 launch — AgentGuard is a pre-deployment guardrail review.
People also search
Click a tab to expand it; click again to hide. All content stays on the page for search engines.
Start smarter today
Join builders using AgentGuard. Free to try — no card needed.
Trusted · Cancel anytime