Cloud API Proxies
Centralized proxying creates one high-impact failure point.
- - Sensitive prompts transit external proxy infrastructure.
- - A proxy outage can stall every downstream request.
- - Control logic and latency budgets become tightly coupled.
Available now - Public beta
7 layers between your AI agents and disaster.
Security decisions made by rules, not predictions. Runs locally. No LLM in the security path. Your data stays in your infrastructure.
1,560+ detection patterns · 7 defense layers · 0 LLMs in the security path
The threat landscape
Real CVEs, active exploit patterns, and evidence from production systems.
The threat landscape is real
These CVEs are active in tools used by millions of developers. Each one was observed in real-world environments.
CVE-2026-25253
OpenClaw token exfiltration
Token exfiltration via malicious gatewayUrl override
CVE-2025-6514
CVSS 9.6
mcp-remote command injection
OS command injection when connecting to untrusted servers
CVE-2025-68143
mcp-server-git path traversal
Path traversal via attacker-controlled paths to Git operations
CVE-2025-68144
mcp-server-git argument injection
Argument injection when tool inputs become part of a Git command line
How we're different
Not another proxy and not just an output filter. GuardClaw enforces layered controls before high-impact actions run.
Output Validators
Checks after generation miss threats that execute earlier.
- - Risk reaches the model before enforcement decisions.
- - Post-hoc checks cannot prevent every side effect.
- - Incident response starts after damage windows open.
GuardClaw
Deterministic controls run before high-impact actions.
- - Runs in your infrastructure with explicit trust boundaries.
- - 7 independent layers gate risky actions before execution.
- - Audit-ready logs preserve clear evidence for every decision.
Explore all 7 layers
GuardClaw enforces deterministic controls before the model executes side effects. That design keeps trust boundaries explicit and incident response faster.
7 layers. Defense in depth.
Each layer operates independently. An attacker must defeat all seven to compromise an agent.
1 Threat Intelligence
CVE and IOC pattern matching, crowd-sourced live threat feeds, and configurable blocklists. Users report new attack vectors; approved patterns protect everyone in real time.
2 Input Validation
Prompt injection detection for tool-call inputs, plus an extensible validation library covering PII redaction, URL/SSRF, path traversal, and command injection. 1,560+ compiled patterns across 11 attack categories.
3 Policy Enforcement
Deny-by-default YAML policies. Allow, deny, or require approval based on tool, action, resource, actor, provider, and untrusted-source context.
4 Capability Tokens
Short-lived, cryptographically signed tokens for every approved action. Single-use, time-bound, scope-limited. HMAC-SHA256 signed and bound to request digest.
5 Sandboxed Execution
Shell wrapper deny-by-default rules, filesystem sandboxing, and HTTP allowlists with SSRF and DNS-rebinding protections.
6 Human-in-the-Loop
High-risk operations pause for human approval. Approvals are bound to request digests to prevent TOCTOU attacks. Optional webhooks for notifications.
7 Receipt Chain
Cryptographically linked audit trail. Every decision recorded with SHA-256 hash chains. Tamper-evident. Built for compliance and forensics.
Try to break it
Launch simulated attacks and watch GuardClaw's seven layers respond in real time.
Attack Types
Defense Layers
Layer 1
Threat Intelligence
Layer 2
Input Validation
Layer 3
Policy Enforcement
Layer 4
Capability Tokens
Layer 5
Sandboxed Execution
Layer 6
Human-in-the-Loop
Layer 7
Receipt Chain
Select an attack type to start
Start in 3 steps
1 Install
brew install guardclaw/tap/guardclaw 2 Initialize
guardclaw init 3 Protect
guardclaw wrap -- your-agent-commandWorks with your stack
Claude MCP
MCP stdio server
Cursor
Shell wrapper
OpenClaw
Guard agent plugin
Docker
Sidecar container
Kubernetes
DaemonSet / sidecar
Cloud Run
Container deployment
Free during public beta
All 7 layers. All 1,560+ patterns. Paid tiers coming after beta.
Available now
Free
during public beta
Everything you need to start securing your AI agents today.
- All 7 defense layers
- 1,560+ detection patterns
- Local-first CLI
- Policy templates
- 1 workspace
- 2 environments
- 2 agents per environment
- 1,000 receipts / month
- 7-day data access
Coming soon
Pro
$49 / month
For teams running multiple agents in production.
- 5 workspaces, 5 environments, 5 agents
- 100,000 receipts / month
- 90-day data access
- Policy packs and receipt export
- Slack and email alerts
- Priority support
Coming soon
Ultimate
$99 / month
Full control for organizations with compliance needs.
- 10 workspaces, 10 environments, 10 agents
- 500,000 receipts / month
- 365-day data access
- Approval workflows and webhooks
- SSO / SAML
- Audit API and custom policies
Need enterprise features now? Let's talk.
Free during public beta. No catches.
All 7 defense layers, 1,560+ patterns, and 0 LLMs in the security path. Install in 3 minutes. No credit card. No restrictions.