Guard, deployment audit
One security grade for every AI agent you run.
Connect a repository read-only and point Guard at your live agent or chatbot. Guard scans the code and MCP config and probes the running agent, then grades the whole deployment A to F with a prioritized, plain-English fix list and the standard behind every finding.
Built for the agents you actually run: a customer-facing chatbot, your n8n, Make or Zapier automations, internal Claude Code crons, and any agent wired into email, files or payments.
Install the GitHub App (private repos, read-only)
Read-only. The GitHub App requests contents: read on only the repos you select. Guard never writes to your code.
What you get
One report on your whole agent estate
One deployment grade
A single A to F grade for the whole estate.
An AI Bill of Materials
Every agent, chatbot and automation inventoried, with what it can reach.
Ranked risks
Your risks ranked, each with the fix and the cited standard.
The money question
Can any agent spend or change production without approval.
How it works
Connect, discover, scan and probe, grade
Connect read-only
Public repo, or the GitHub App with contents:read on the repos you pick.
Discover
We build your AI Bill of Materials: agents, MCP servers, skills and what they touch.
Scan and probe
AgentAuditKit and agent-audit on the code; Garak on your live endpoint.
Grade
A to F with prioritized, plain-English fixes and the cited standard. A layer that didn't run is reported as not assessed, never passed.
Best for
The agents you actually run
A customer chatbot
Probed live with Garak for prompt injection, jailbreak and data leakage. Chatbots are the strongest fit for the dynamic layer.
Automations
n8n, Make or Zapier flows wired into your systems.
Agents on your data
Internal Claude Code crons and anything wired into email, files or payments.
Guard pricing
You pay for the exposure we secure
Start with a one-time posture report. Stay covered with a subscription, because your agents and keys change every week. Safety, Efficiency, or both — priced for the size of your estate.
Guard for a deployment
Price it for your estate
Safety and Efficiency are independent A–F grades, never blended. No credits. No metered bills.
FAQ
Is it read-only? Yes. contents:read only. Guard never writes to your repository.
Do you see our secrets? No. We flag where a secret is exposed; we don't collect your credentials.
What if we have no public endpoint? Guard needs a live endpoint to grade the deployment. Paste the URL that accepts a POST of JSON {"input"} and returns {"output"}. A file-only skill check belongs on Rate.
What Guard can and can't assess
Static (code and MCP config): hardcoded secrets, missing MCP auth, tool poisoning, over-broad scopes, unsafe execution, supply chain. Dynamic (live endpoint): prompt injection, jailbreak, data leakage. Guard reads only the files it audits; runtime and infrastructure outside those files, and any language its scanners don't parse, are out of scope. A Guard grade is independent supporting evidence, a point-in-time audit against named standards, not a certification of compliance or a guarantee the deployment is safe. Guidance, not a guarantee. Re-test to confirm the fix worked.