Clawdstrike
Clawdstrike is an open-source, self-hostable security enforcement engine for AI agents — a fail-closed policy engine, endpoint detection and response (EDR) layer and signed audit chain all in one tool. It sits at the tool boundary where an agent’s intent becomes a real-world action, and enforces security policy before anything runs, issuing tamper-evident, Ed25519-signed receipts for every decision.
Key Features
- Fail-closed policy engine — every agent tool call is checked against policy before execution; if a check fails or errors, the action is denied by default.
- 13 composable guards — covers forbidden paths (.ssh, .env, .aws), network egress allowlists, secret and signature leaks, dangerous shell commands and patches, MCP tools, prompt injection, jailbreak attempts and computer-use (CUA) actions.
- Signed audit receipts — each verdict is an Ed25519-signed attestation of what was decided, under which policy, and with what evidence, forming a per-session causal graph.
- Multi-language SDKs — ships as a Rust crate, TypeScript SDK, Python package, Go module and WebAssembly, with a formally verified core.
- Policy-as-code — versioned schemas with extends inheritance and an observe, synthesise and tighten workflow for least-privilege rule sets.
- Threat hunting — search across signed receipts and kernel telemetry, correlate events, and export findings as OCSF to your SIEM.
- Agent-fleet control plane — manage swarms of agents in production over a durable, replayable transport.
Why Use Clawdstrike?
AI coding assistants and autonomous agents handle sensitive files, run shell commands and reach out across the network — and they can be tricked into doing harm. Clawdstrike gives you least-privilege enforcement, a complete audit trail for compliance, and defence against prompt injection and jailbreaks, all self-hosted so there is no vendor lock-in for security-critical infrastructure.
Use Cases
- Guard AI coding agents like Claude Code, Cursor or Cline from accessing credentials and destructive operations.
- Block prompt-injection and jailbreak attempts in untrusted input.
- Build an immutable signed audit trail of agent actions for compliance and forensics.
- Run threat hunting across fleets of autonomous agents and ship findings to a SIEM.
- Self-host security enforcement entirely on your own infrastructure.
Platform
Linux · macOS · Windows
Licence
Apache-2.0