7 layers between your AI agents and disaster.

GuardClaw is a deterministic runtime security layer for AI agents. It enforces 7 independent defense layers — from threat intelligence to receipt chains — using policy rules, not probabilistic LLM inference. It runs locally, keeps data in your infrastructure, and makes every security decision auditable and repeatable.

Most guardrail products use a second LLM to judge the first. That means your security path is probabilistic, slow, and opaque. GuardClaw uses deterministic policy rules — pattern matching, allowlists, scope locks, and cryptographic receipt chains — so every decision is fast, repeatable, and auditable. No LLM sits in the security path.

No. GuardClaw deliberately keeps AI out of the security decision path. All enforcement is deterministic: policy rules, pattern matching, and scope validation. This makes security outcomes repeatable and auditable, unlike probabilistic LLM-based approaches that can be bypassed with prompt injection.

GuardClaw defends against prompt injection (direct and indirect), tool misuse, scope escalation, data exfiltration, unauthorized actions, and supply-chain attacks on agent toolchains. Its 7 layers are: Threat Intelligence, Input Validation, Policy Enforcement, Capability Tokens, Sandboxed Execution, Human-in-the-Loop, and Receipt Chain.

Install GuardClaw via Homebrew (macOS/Linux): brew install TakeInterestInc/tap/guardclaw. Then run guardclaw doctor to check your setup and guardclaw test --audit to score your agent configuration. The full Getting Started tutorial walks you through everything in five minutes: takeinterest.ai/blog/getting-started-with-guardclaw/

GuardClaw is proprietary software licensed under the GuardClaw Proprietary License (EULA). It is not open source. The core runtime runs entirely in your infrastructure with no external dependencies. The Free tier is available at no cost during and after public beta, with Pro and Ultimate paid tiers for teams that need higher quotas and advanced features.

GuardClaw runs locally by default. All policy evaluation, threat detection, and enforcement happen in your infrastructure. Anonymous telemetry (decision counts, threat scores, timing, platform, version) is enabled by default to improve security patterns for the community and can be disabled at any time with guardclaw telemetry disable. No raw prompts, commands, file paths, or PII are ever collected. On paid plans (Pro and Ultimate), you can also opt out of anonymized training data collection.

Yes. GuardClaw supports multiple strictness levels — paranoid, strict, balanced, and permissive — so you can tune enforcement to your workflow. Paranoid blocks everything not explicitly allowed. Permissive only blocks known-dangerous patterns. You can also write custom allow and deny rules in YAML policies for fine-grained control.

Yes. GuardClaw's Threat Intelligence layer includes patterns for known CVEs affecting AI agent toolchains, along with live threat feeds. When a new vulnerability is reported and verified, detection patterns are distributed to all authenticated users automatically.

GuardClaw includes self-reference protection — detection patterns that identify attempts by AI agents to disable, bypass, or modify GuardClaw itself. If an agent tries to uninstall GuardClaw, modify its configuration, or suppress its hooks, those actions are blocked and logged.